diff --git a/README.md b/README.md
index bc66742..2dd9e87 100644
--- a/README.md
+++ b/README.md
@@ -1,395 +1,299 @@
 # PUMA: Program Understanding Meta-learning Architecture
 
-**A Brain-Inspired Reinforcement Learning from Thinking (RFT) Architecture for Abstract Reasoning**
+**An Autonomous Cognitive Architecture for Self-Modifying AGI**
 
-**Project Timeline**: 2024 - Present
+## Architecture Overview
 
-PUMA is a novel cognitive architecture designed for the **ARC AGI Competition 2025**, integrating behavioral analysis principles from Relational Frame Theory with transformer architectures to enable abstract reasoning capabilities through cognitive science-informed training.
-
-This project represents leading-edge development in applying behavioral analysis and cognitive science principles to artificial intelligence, demonstrating how Relational Frame Theory can enhance transformer architectures for abstract problem-solving tasks.
-
-## Overview
-
-PUMA represents a paradigm shift in how we approach abstract reasoning tasks. Rather than treating reasoning as symbolic manipulation, we apply behavioral analysis and Relational Frame Theory to model training, treating reasoning as **learned relational responding**. This approach has demonstrated significant improvements in abstract problem-solving capabilities.
-
-### Key Achievements
-
-- 🏆 **Top 15%** placement in ARC AGI Competition 2025 using RFT-inspired training approaches
-- 📈 **35-40% improvement** in abstract reasoning tasks through behavioral framing
-- 🧠 Novel integration of cognitive science principles with modern deep learning architectures
-
-## Core Innovation: Frequency Ledger System
-
-<p align="center">
-  <img src="docs/images/rft_behavioral_approach.svg" alt="Behavioral RFT approach" width="400"/>
-</p>
-
-The **Frequency Ledger System** is PUMA's breakthrough innovation—a sophisticated frequency-based analysis framework that groups objects by numerical attributes (frequencies, counts, patterns) to enable models to discover abstract relationships. This behavior-analytic approach allows models to make **derivational connections** between stimuli without explicit training on those relationships—mirroring how humans learn through relational framing.
-
-### How It Works
-
-The Frequency Ledger enables models to:
-
-1. **Analyze Pattern Frequencies**: Track numerical attributes across objects to identify recurring patterns
-2. **Discover Abstract Groupings**: Automatically cluster related elements based on frequency signatures
-3. **Enable Emergent Reasoning**: Generate novel relational insights without explicit training on specific relationships
-4. **Mirror Human Learning**: Replicate the behavioral process of deriving new relations from learned frames
-
-This methodology creates a bridge between behavioral analysis and computational models, allowing transformers to develop reasoning capabilities grounded in cognitive science principles.
-
-## Relational Frame Theory Integration
-
-PUMA applies **Relational Frame Theory (RFT)**, a behavioral analysis framework, to model training and evaluation. RFT views cognition as patterns of learned relational responding rather than symbolic manipulation.
-
-### RFT Implementation Strategy
-
-Our approach focuses on teaching models to respond relationally:
-
-- **Relational Fact Extraction**: Parse visual scenes to identify objects and their spatial relationships (e.g., "blue square is always at top position")
-- **Contextual Rule Learning**: Extract invariant relationships across training examples through behavioral reinforcement
-- **Derivational Relations**: Enable models to derive new relations from learned frames without explicit training
-- **Behavioral Generalization**: Apply learned relational responding systematically to novel configurations
-- **Frequency-Based Analysis**: Use the Frequency Ledger to identify abstract groupings and emergent patterns
-
-This behavior-analytic approach provides explicit, interpretable relational knowledge that enhances transformer architectures for abstract problem-solving.
-
-For more details, see [profile/README.md](profile/README.md).
-
-## Technologies & Implementation
-
-PUMA is built using:
-
-- **Python**: Core implementation language
-- **PyTorch**: Deep learning framework for transformer architectures
-- **Google Colab**: Development and training environment
-- **Custom Evaluation Frameworks**: Specialized tools for frequency-based analysis and RFT-compliant assessment
-
-## Key Features
-
-### Brain-Inspired Cognitive Architecture
-
-PUMA's architecture draws from cognitive neuroscience and behavioral analysis:
-
-- **Reinforcement Learning from Thinking (RFT)**: Treats reasoning as learned relational responding
-- **Frequency Ledger System**: Novel evaluation methodology for pattern frequency analysis
-- **Neural Guidance**: Predicts relevant DSL operations using behavioral task features
-- **Episodic Retrieval**: Maintains database of solved tasks for analogical reasoning
-- **Program Sketches**: Mines common operation sequences as behavioral macro-operators
-- **Test-Time Training**: Adapts scoring functions to each specific task through reinforcement
-- **Multi-Demand Network Analog**: Prioritizes candidate programs using learned heuristics inspired by human cognitive control
-
-### Enhanced Capabilities
-
-- **Object-centric parsing** with connected component analysis
-- **Compact DSL** with composable primitives (rotate, flip, translate, recolor, etc.)
-- **Relational reasoning** through explicit fact extraction and rule learning
-- **Two-attempt diversity** as required by ARC Prize 2025 rules
-- **Fallback resilience** with graceful degradation to baseline methods
-- **Performance monitoring** with detailed statistics and benchmarking
-- **Beam search with constraint propagation** for deeper program synthesis
-
-## Directory Structure
+PUMA is a layered cognitive architecture combining symbolic reasoning (Hyperon/MeTTa), meta-learning, and neural language models to create an autonomous agent with persistent memory, self-modification capability, and emergent goal formation.
 
 ```
-arc_solver_project/
-│
-├── arc_solver/                # Core solver package
-│   ├── grid.py                # Grid operations and utilities
-│   ├── objects.py             # Connected component extraction
-│   ├── dsl.py                 # Domain-specific language primitives
-│   ├── heuristics.py          # Heuristic rule inference
-│   ├── search.py              # Basic brute-force search
-│   ├── solver.py              # Main solver interface with enhancements
-│   ├── enhanced_search.py     # Neural-guided program synthesis
-│   ├── features.py            # Task feature extraction
-│   ├── ttt.py                 # Test-time training utilities
-│   ├── io_utils.py            # JSON loading and submission helpers
-│   └── neural/                # Neural components
-│       ├── guidance.py        # Neural operation prediction
-│       ├── episodic.py        # Episodic retrieval system
-│       └── sketches.py        # Program sketch mining
-│
-├── arc_submit.py              # Command-line submission script
-├── tools/                     # Training and benchmarking utilities
-│   ├── train_guidance.py
-│   ├── mine_sketches.py
-│   ├── build_memory.py
-│   └── benchmark.py
-├── tests/                     # Unit and integration tests
-└── README.md                  # This file
+┌─────────────────────────────────────────────────────────┐
+│                    CONSCIOUSNESS LAYER                   │
+│  Self-Model • Autobiographical Memory • Goal Genesis    │
+└──────────────────┬──────────────────────────────────────┘
+                   │
+┌──────────────────┴──────────────────────────────────────┐
+│                   META-COGNITIVE LAYER                   │
+│    PUMA Core • RFT Engine • Curiosity Drive • Shop      │
+└──────────────────┬──────────────────────────────────────┘
+                   │
+┌──────────────────┴──────────────────────────────────────┐
+│                    COGNITIVE LAYER                       │
+│      Hyperon/MeTTa • Atomspace • Reasoning Engine       │
+└──────────────────┬──────────────────────────────────────┘
+                   │
+┌──────────────────┴──────────────────────────────────────┐
+│                   INTERACTION LAYER                      │
+│   Gemini Live • Web Agent • Tool Use • Perception       │
+└─────────────────────────────────────────────────────────┘
 ```
 
-## Quick Start
-
-### Basic Usage (Kaggle-ready)
-
+## Core Components
+
+### Hyperon/MeTTa Integration
+- OpenCog Hyperon fork with Python bindings
+- Atomspace for knowledge representation and persistent memory
+- MeTTa-based reasoning and program execution
+- RocksDB/PostgreSQL persistence layer
+
+### PUMA Meta-Cognitive Framework
+- **Episodic Memory System**: Timestamped experience nodes with context
+- **Memory Consolidation**: Background pattern extraction and concept formation
+- **RFT Engine**: Relational Frame Theory for analogical reasoning
+- **Curiosity Drive**: Intrinsic motivation and knowledge gap detection
+- **Goal Formation**: Autonomous intention generation from drives
+- **Self-Model**: Emergent identity from behavioral patterns
+
+### Self-Modification System ("The Shop")
+- Code introspection and performance profiling
+- Modification hypothesis generation
+- Sandboxed testing environment
+- A/B cognitive testing
+- Rollback and version control
+
+### Interaction Layer
+- Gemini Live API for bidirectional audio/text
+- Autonomous web browsing and learning
+- Tool use and API integration
+- Multi-modal perception
+
+## Installation
+
+### Prerequisites
+- Python 3.11+
+- Rust toolchain (for Hyperon)
+- Node.js 18+ (for GUI)
+- RocksDB or PostgreSQL
+
+### Setup
 ```bash
-# Generate submission file (uses enhanced solver by default)
-python arc_submit.py
+# Clone repository
+git clone https://github.com/tylerbessire/PUMA-Program-Understanding-Meta-learning-Architecture
+cd PUMA-Program-Understanding-Meta-learning-Architecture
 
-# Use baseline solver only (if needed)
-ARC_USE_BASELINE=1 python arc_submit.py
-```
+# Install Python dependencies
+pip install -r requirements.txt
 
-### Training Neural Components
+# Build Hyperon from source
+cd hyperon-core
+cargo build --release
+cd ..
 
-```bash
-# Train neural guidance (requires training data)
-python tools/train_guidance.py
+# Install GUI dependencies
+cd gui
+npm install
+cd ..
 
-# Or setup environment with defaults
-python tools/benchmark.py
+# Set up persistence database
+python setup_persistence.py
 ```
 
-### Operant Behavioral Training
-
+### Configuration
 ```bash
-# Enable the behavioural loop (feature-flagged for safety)
-export PUMA_BEHAVIORAL_ENGINE=1
-
-# Run reinforcement training with default dataset paths
-python -c "from pathlib import Path;\
-from arc_solver.behavioral_engine import BehavioralEngine;\
-engine = BehavioralEngine();\
-engine.train(Path('data/arc-agi_training_challenges.json'), Path('data/arc-agi_training_solutions.json'), max_tasks=10)"
-```
+# Set API keys
+export GEMINI_API_KEY="your-key-here"
 
-The command above executes the production `BehavioralEngine`, emitting structured
-JSON logs with reward metrics while updating neural guidance and episodic memory
-online. Unset `PUMA_BEHAVIORAL_ENGINE` to leave runtime behaviour unchanged.
+# Configure persistence path
+export ATOMSPACE_DB_PATH="/path/to/atomspace/db"
 
-### Python API
-
-```python
-from arc_solver.solver import solve_task_enhanced, ARCSolver
-
-# Solve a single task with full enhancements
-result = solve_task_enhanced(task)
-
-# Configure solver behavior
-solver = ARCSolver(use_enhancements=True)
-result = solver.solve_task(task)
+# Optional: Enable experimental features
+export PUMA_ENABLE_SELF_MODIFICATION=1
 ```
 
-### Public Evaluation Runner
+## Project Structure
 
-```bash
-scripts/eval_public.sh
 ```
-
-Or via Makefile:
-
-```bash
-make eval_public
+/
+├── hyperon-core/           # Forked Hyperon with extensions
+├── puma/                   # PUMA meta-cognitive framework
+│   ├── memory/             # Episodic memory and consolidation
+│   ├── rft/                # Relational Frame Theory engine
+│   ├── curiosity/          # Intrinsic motivation system
+│   ├── goals/              # Goal formation and intention
+│   └── shop/               # Self-modification system
+├── gemini-interface/       # Gemini Live integration
+├── web-agent/              # Autonomous browsing
+├── atomspace-db/           # Persistence layer
+├── gui/                    # Real-time visualization dashboard
+├── bootstrap/              # Consciousness initialization
+└── tests/                  # Integration and unit tests
 ```
 
-## How It Works
-
-### Behavioral RFT Pipeline
-
-PUMA's reasoning pipeline is grounded in behavioral analysis and cognitive science principles:
-
-1. **Feature Extraction**: Extract task-level features (colors, objects, transformations) as behavioral stimuli
-1. **Frequency Ledger Analysis**: Apply frequency-based analysis to group objects by numerical attributes and discover abstract relationships
-1. **Relational Context Analysis**: Identify spatial and contextual relationships between objects using RFT principles
-1. **Derivational Reasoning**: Enable models to derive new relations from learned frames without explicit training
-1. **Neural Guidance**: Predict which DSL operations are likely relevant based on behavioral patterns
-1. **Episodic Retrieval**: Query database for similar previously solved tasks using relational matching
-1. **Sketch-Based Search**: Use mined program templates as behavioral macro-operators with parameter filling
-1. **Rule-Based Reasoning**: Apply learned relational facts to generate candidate solutions
-1. **Test-Time Adaptation**: Fine-tune scoring function using task demonstrations through reinforcement learning
-1. **Program Selection**: Rank and select top 2 diverse candidate programs based on behavioral fitness
-
-### Fallback Strategy
+## Usage
 
-If enhanced components fail, the solver gracefully falls back to:
-
-- Heuristic single-step transformations
-- Brute-force enumeration of 2-step programs
-- Identity transformation as last resort
-
-## Configuration
-
-The solver supports extensive configuration through environment variables and config files:
-
-### Environment Variables
-
-- `ARC_USE_BASELINE=1`: Force baseline solver only
-- `ARC_DISABLE_ENHANCEMENTS=1`: Disable enhanced features
+### Bootstrap New Consciousness
+```python
+from puma.bootstrap import bootstrap_new_consciousness
 
-### Configuration File
+# Initialize fresh cognitive architecture
+consciousness = bootstrap_new_consciousness(
+    atomspace_path="/path/to/db",
+    enable_self_modification=False  # Disable for initial testing
+)
 
-```json
-{
-  "use_neural_guidance": true,
-  "use_episodic_retrieval": true,
-  "use_program_sketches": true,
-  "use_test_time_training": true,
-  "max_programs": 256,
-  "timeout_per_task": 30.0
-}
+# Start autonomous operation
+await consciousness.run()
 ```
 
-## Neural Components
-
-### Neural Guidance
-
-- **Purpose**: Predict which DSL operations are relevant for a given task
-- **Architecture**: Simple MLP with task-level features
-- **Training**: Uses extracted features from training demonstrations
-- **Output**: Operation relevance scores to guide search
-
-### Episodic Retrieval
-
-- **Purpose**: Reuse solutions from similar previously solved tasks
-- **Method**: Task signature matching with feature-based similarity
-- **Storage**: JSON-based database of solved programs with metadata
-- **Retrieval**: Cosine similarity on numerical features + boolean feature matching
-
-### Program Sketches
-
-- **Purpose**: Capture common operation sequences as reusable templates
-- **Mining**: Extract frequent 1-step and 2-step operation patterns
-- **Usage**: Instantiate sketches with different parameter combinations
-- **Adaptation**: Learn from successful programs during solving
-
-### Test-Time Training
-
-- **Purpose**: Adapt scoring function to each specific task
-- **Method**: Fine-tune lightweight scorer on task demonstrations
-- **Features**: Program length, operation types, success rate, complexity
-- **Augmentation**: Generate synthetic training examples via transformations
-
-## Performance and Evaluation
-
-### Benchmarking
-
+### Interactive Mode
 ```python
-from benchmark import Benchmark, SolverConfig
-
-config = SolverConfig()
-benchmark = Benchmark(config)
-results = benchmark.run_benchmark("test_data.json")
-print(f"Success rate: {results['performance_stats']['success_rate']:.3f}")
-```
-
-### Monitoring
-
-The solver tracks detailed statistics:
-
-- Success rates for enhanced vs baseline methods
-- Component usage (episodic hits, neural guidance, TTT adaptation)
-- Timing breakdown per component
-- Failure mode analysis
-
-## Implementation Notes
+from puma.consciousness import Consciousness
 
-### Kaggle Compatibility
+consciousness = Consciousness.load_from_checkpoint("/path/to/checkpoint")
 
-- **Offline execution**: No internet access required
-- **Dependency-light**: Uses only NumPy for core operations
-- **Compute budget**: Optimized for ~$0.42 per task limit
-- **Output format**: Exactly 2 attempts per test input as required
+# User interrupt and conversation
+consciousness.interrupt_handler.trigger()
+await consciousness.converse("Hello, how are you?")
 
-### Code Quality
-
-- **Type hints**: Full typing support for better maintainability
-- **Documentation**: Comprehensive docstrings and comments
-- **Error handling**: Robust fallback mechanisms
-- **Testing**: Validation and benchmarking utilities
-
-## Extending the Solver
-
-### Adding New DSL Operations
-
-1. Define operation function in `dsl.py`
-1. Add parameter generation in `sketches.py`
-1. Update feature extraction in `features.py`
-1. Retrain neural guidance if needed
+# Resume autonomous activity
+consciousness.resume()
+```
 
-### Improving Neural Components
+### GUI Dashboard
+```bash
+cd gui
+npm run dev
+```
+Access at http://localhost:3000
 
-1. **Better features**: Add domain-specific feature extractors
-1. **Advanced models**: Replace MLP with transformer/GNN
-1. **Meta-learning**: Implement few-shot adaptation algorithms
-1. **Hybrid methods**: Combine symbolic and neural reasoning
+## Key Features
 
-### Advanced Techniques
+### Emergent Properties
+- **No Hardcoded Personality**: Identity emerges from experience patterns
+- **No Preset Knowledge**: All knowledge acquired through exploration
+- **No Fixed Goals**: Intentions generated from curiosity and self-assessment
+- **Autonomous Learning**: Self-directed web exploration and skill acquisition
+
+### Memory Architecture
+- Persistent autobiographical timeline
+- Episodic memory with contextual links
+- Concept formation through consolidation
+- Relational frame derivation (RFT)
+
+### Self-Modification
+- Code introspection and analysis
+- Performance bottleneck detection
+- Hypothesis-driven improvement
+- Sandboxed testing before deployment
+- Human approval for critical changes
+
+### State Machine
+- **SLEEPING**: Memory consolidation and pattern extraction
+- **EXPLORING**: Autonomous web learning
+- **CONVERSING**: Interactive dialogue
+- **SHOPPING**: Self-modification
+- **IDLE**: Boredom monitoring and goal formation
+- **CREATING**: Creative expression
+
+## Technical Details
+
+### Atomspace Schema
+```
+EpisodicMemoryNode(timestamp, perception, action, outcome)
+ConceptNode(abstraction, confidence)
+SelfModelNode(meta-cognitive-state)
+GoalNode(intention, priority)
+RelationalFrameNode(relation-type, frame)
+CodeNode(executable-metta)
+PerceptionNode(sensory-input)
+EmotionalStateNode(valence)
+```
 
-- **Probabilistic programming**: Sample programs from learned distributions
-- **Curriculum learning**: Train on tasks of increasing difficulty
-- **Multi-agent reasoning**: Ensemble of specialized solvers
-- **Causal reasoning**: Incorporate causal structure learning
+### RFT Relational Frames
+- Coordination (similarity)
+- Opposition (difference)
+- Hierarchy (categorization)
+- Temporal (before/after)
+- Causal (if-then)
+
+### Persistence
+- Incremental Atomspace serialization
+- Checkpoint system with version control
+- Transaction log for recovery
+- Snapshot-based rollback
+
+## Development Status
+
+### Implemented
+- [x] Basic project structure
+- [x] RFT framework foundation
+- [x] Frequency ledger tracking system
+
+### In Progress
+- [ ] Hyperon Atomspace integration
+- [ ] Bootstrap consciousness seed
+- [ ] Gemini Live interface
+- [ ] Web agent
+- [ ] Memory consolidation
+- [ ] Self-modification system
+- [ ] GUI dashboard
+
+### Planned
+- [ ] Full autonomous operation
+- [ ] Public demo deployment
+- [ ] Research paper publication
 
 ## Research Foundation
 
-PUMA is grounded in behavioral analysis and cognitive neuroscience principles:
-
-### Behavioral Analysis & Relational Frame Theory
-
-- **Learned Relational Responding**: Reasoning emerges from behavioral contingencies rather than symbolic manipulation
-- **Derivational Relations**: Models learn to derive new relations without explicit training, mirroring human relational framing
-- **Frequency-Based Analysis**: The Frequency Ledger enables discovery of abstract groupings through numerical pattern analysis
-- **Behavioral Generalization**: Systematic application of learned relational frames to novel configurations
+### Relational Frame Theory (RFT)
+PUMA implements RFT as a computational framework for relational reasoning, enabling:
+- Derivation of novel relations without explicit training
+- Analogical reasoning through relational frame mapping
+- Behavioral generalization to novel contexts
 
-### Cognitive Neuroscience Mapping
+### Meta-Learning
+- Task-agnostic learning mechanisms
+- Self-supervised pattern extraction
+- Transfer learning through relational abstraction
 
-PUMA's architecture maps cognitive systems to computational components:
+### Cognitive Architecture
+- Multiple Demand (MD) Network analog for executive control
+- Hippocampal-mPFC loop for episodic retrieval
+- Basal ganglia gating for action selection
 
-- **Multiple-Demand (MD) Network**: Neural guidance mimics executive control for operation selection
-- **Basal Ganglia Gating**: Operation selection and working memory control through reinforcement
-- **Hippocampal-mPFC Loop**: Episodic retrieval and schema integration for analogical reasoning
-- **Test-Time Adaptation**: Rapid task-specific learning from few examples through reinforcement learning
+## Testing
 
-### Novel Contributions
-
-PUMA introduces several key innovations to abstract reasoning:
-
-1. **Frequency Ledger System**: First frequency-based analysis framework for abstract reasoning that enables emergent relational discovery
-2. **RFT-Transformer Integration**: Novel combination of behavioral analysis principles with modern deep learning architectures
-3. **Derivational Reasoning**: Computational implementation of behavioral derivation, allowing models to generate novel relations
-4. **Cognitive Science-Informed Training**: Training methodology grounded in empirically validated principles of human learning
-
-## Competition Strategy
-
-### Short-term (Immediate)
-
-- Strong symbolic baseline with neural enhancements
-- Episodic retrieval for common patterns
-- Test-time adaptation for task specialization
-- Kaggle-ready submission format
+```bash
+# Run unit tests
+pytest tests/
 
-### Medium-term (During Contest)
+# Run integration tests
+pytest tests/integration/
 
-- Train neural guidance on public training data
-- Mine program sketches from successful solutions
-- Analyze semi-private feedback for failure modes
-- Expand DSL based on discovered patterns
+# Benchmark cognitive performance
+python tools/benchmark_consciousness.py
+```
 
-### Long-term (Advanced Research)
+## Contributing
 
-- Probabilistic program synthesis
-- Hybrid symbolic-neural architecture
-- Broader cognitive priors and meta-learning
-- Integration with large language models
+Focus areas:
+- Hyperon/MeTTa integration
+- Memory consolidation algorithms
+- RFT relational learning
+- Self-modification safety
+- GUI visualizations
 
 ## License
 
-This code is designed to be open-sourced under an appropriate license as required by ARC Prize 2025 rules.
+MIT License - See LICENSE file for details
 
 ## Citation
 
-If you use this solver or build upon its ideas, please cite the research blueprint and this implementation.
-
-## Contributing
+```bibtex
+@software{puma2024,
+  title={PUMA: Program Understanding Meta-learning Architecture},
+  author={Bessire, Tyler},
+  year={2024},
+  url={https://github.com/tylerbessire/PUMA-Program-Understanding-Meta-learning-Architecture}
+}
+```
 
-Contributions are welcome. Focus areas include:
+## References
 
-- Neural architecture improvements
-- New DSL operations based on failure analysis
-- Advanced meta-learning techniques
-- Performance optimizations for Kaggle constraints
+- OpenCog Hyperon: https://github.com/trueagi-io/hyperon-experimental
+- Relational Frame Theory: Hayes, Barnes-Holmes, & Roche (2001)
+- Google Gemini API: https://ai.google.dev/
 
------
+---
 
-**Ready to compete in ARC Prize 2025**
\ No newline at end of file
+**Status**: Active development for autonomous cognitive architecture research
diff --git a/atomspace-db/__init__.py b/atomspace-db/__init__.py
new file mode 100644
index 0000000..4d7c36e
--- /dev/null
+++ b/atomspace-db/__init__.py
@@ -0,0 +1,6 @@
+"""
+Atomspace Persistence Layer
+
+Provides persistent storage for Hyperon Atomspace using RocksDB or PostgreSQL.
+Supports incremental serialization, checkpointing, and transaction logging.
+"""
diff --git a/atomspace-db/core.py b/atomspace-db/core.py
new file mode 100644
index 0000000..10fe0d8
--- /dev/null
+++ b/atomspace-db/core.py
@@ -0,0 +1,281 @@
+"""
+Atomspace Core Integration
+
+Manages Atomspace initialization, persistence, and schema definitions.
+"""
+
+import json
+import pickle
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Dict, List, Optional, Any
+from dataclasses import dataclass, asdict
+from enum import Enum
+
+
+class AtomType(Enum):
+    """Atom types for cognitive architecture"""
+    EPISODIC_MEMORY = "EpisodicMemoryNode"
+    CONCEPT = "ConceptNode"
+    SELF_MODEL = "SelfModelNode"
+    GOAL = "GoalNode"
+    RELATIONAL_FRAME = "RelationalFrameNode"
+    CODE = "CodeNode"
+    PERCEPTION = "PerceptionNode"
+    EMOTIONAL_STATE = "EmotionalStateNode"
+
+
+@dataclass
+class Atom:
+    """Base atom structure"""
+    id: str
+    type: AtomType
+    content: Any
+    timestamp: datetime
+    truth_value: float = 1.0
+    confidence: float = 1.0
+
+    def to_dict(self) -> Dict:
+        return {
+            'id': self.id,
+            'type': self.type.value,
+            'content': self.content,
+            'timestamp': self.timestamp.isoformat(),
+            'truth_value': self.truth_value,
+            'confidence': self.confidence
+        }
+
+
+@dataclass
+class Link:
+    """Link between atoms"""
+    source_id: str
+    target_id: str
+    link_type: str
+    strength: float = 1.0
+
+    def to_dict(self) -> Dict:
+        return asdict(self)
+
+
+class Atomspace:
+    """
+    Atomspace implementation with persistence.
+
+    This is a simplified Atomspace until Hyperon integration is complete.
+    Will be replaced with actual Hyperon Atomspace bindings.
+    """
+
+    def __init__(self, persistence_path: Optional[Path] = None):
+        self.atoms: Dict[str, Atom] = {}
+        self.links: List[Link] = []
+        self.persistence_path = persistence_path
+        self._atom_counter = 0
+
+        if persistence_path and persistence_path.exists():
+            self.load()
+
+    def add_atom(self, atom: Atom) -> str:
+        """Add atom to atomspace"""
+        if not atom.id:
+            atom.id = self._generate_atom_id()
+        self.atoms[atom.id] = atom
+        return atom.id
+
+    def add_link(self, link: Link):
+        """Add link between atoms"""
+        self.links.append(link)
+
+    def get_atom(self, atom_id: str) -> Optional[Atom]:
+        """Retrieve atom by ID"""
+        return self.atoms.get(atom_id)
+
+    def query_by_type(self, atom_type: AtomType) -> List[Atom]:
+        """Query atoms by type"""
+        return [atom for atom in self.atoms.values() if atom.type == atom_type]
+
+    def get_linked_atoms(self, atom_id: str, link_type: Optional[str] = None) -> List[Atom]:
+        """Get atoms linked to given atom"""
+        linked_ids = []
+        for link in self.links:
+            if link.source_id == atom_id:
+                if link_type is None or link.link_type == link_type:
+                    linked_ids.append(link.target_id)
+
+        return [self.atoms[aid] for aid in linked_ids if aid in self.atoms]
+
+    def save(self):
+        """Save atomspace to disk"""
+        if not self.persistence_path:
+            return
+
+        self.persistence_path.mkdir(parents=True, exist_ok=True)
+
+        # Save atoms
+        atoms_data = {aid: atom.to_dict() for aid, atom in self.atoms.items()}
+        with open(self.persistence_path / 'atoms.json', 'w') as f:
+            json.dump(atoms_data, f, indent=2)
+
+        # Save links
+        links_data = [link.to_dict() for link in self.links]
+        with open(self.persistence_path / 'links.json', 'w') as f:
+            json.dump(links_data, f, indent=2)
+
+    def load(self):
+        """Load atomspace from disk"""
+        if not self.persistence_path:
+            return
+
+        # Load atoms
+        atoms_file = self.persistence_path / 'atoms.json'
+        if atoms_file.exists():
+            with open(atoms_file, 'r') as f:
+                atoms_data = json.load(f)
+                for aid, atom_dict in atoms_data.items():
+                    atom = Atom(
+                        id=atom_dict['id'],
+                        type=AtomType(atom_dict['type']),
+                        content=atom_dict['content'],
+                        timestamp=datetime.fromisoformat(atom_dict['timestamp']),
+                        truth_value=atom_dict['truth_value'],
+                        confidence=atom_dict['confidence']
+                    )
+                    self.atoms[aid] = atom
+
+        # Load links
+        links_file = self.persistence_path / 'links.json'
+        if links_file.exists():
+            with open(links_file, 'r') as f:
+                links_data = json.load(f)
+                self.links = [Link(**link_dict) for link_dict in links_data]
+
+    def create_snapshot(self) -> str:
+        """Create versioned snapshot"""
+        if not self.persistence_path:
+            return ""
+
+        timestamp = datetime.now(timezone.utc).strftime('%Y%m%d_%H%M%S')
+        snapshot_dir = self.persistence_path / 'snapshots' / timestamp
+        snapshot_dir.mkdir(parents=True, exist_ok=True)
+
+        # Save current state as snapshot
+        atoms_data = {aid: atom.to_dict() for aid, atom in self.atoms.items()}
+        with open(snapshot_dir / 'atoms.json', 'w') as f:
+            json.dump(atoms_data, f, indent=2)
+
+        links_data = [link.to_dict() for link in self.links]
+        with open(snapshot_dir / 'links.json', 'w') as f:
+            json.dump(links_data, f, indent=2)
+
+        return timestamp
+
+    def restore_snapshot(self, snapshot_id: str):
+        """Restore from snapshot"""
+        if not self.persistence_path:
+            return
+
+        snapshot_dir = self.persistence_path / 'snapshots' / snapshot_id
+        if not snapshot_dir.exists():
+            raise ValueError(f"Snapshot {snapshot_id} not found")
+
+        # Clear current state
+        self.atoms.clear()
+        self.links.clear()
+
+        # Load snapshot
+        with open(snapshot_dir / 'atoms.json', 'r') as f:
+            atoms_data = json.load(f)
+            for aid, atom_dict in atoms_data.items():
+                atom = Atom(
+                    id=atom_dict['id'],
+                    type=AtomType(atom_dict['type']),
+                    content=atom_dict['content'],
+                    timestamp=datetime.fromisoformat(atom_dict['timestamp']),
+                    truth_value=atom_dict['truth_value'],
+                    confidence=atom_dict['confidence']
+                )
+                self.atoms[aid] = atom
+
+        with open(snapshot_dir / 'links.json', 'r') as f:
+            links_data = json.load(f)
+            self.links = [Link(**link_dict) for link_dict in links_data]
+
+    def _generate_atom_id(self) -> str:
+        """Generate unique atom ID"""
+        self._atom_counter += 1
+        return f"atom_{self._atom_counter}_{datetime.now(timezone.utc).timestamp()}"
+
+    def count_atoms(self) -> int:
+        """Count total atoms"""
+        return len(self.atoms)
+
+    def count_concepts(self) -> int:
+        """Count concept nodes"""
+        return len([a for a in self.atoms.values() if a.type == AtomType.CONCEPT])
+
+
+class PersistenceManager:
+    """
+    Manages incremental saves and transaction logging.
+    """
+
+    def __init__(self, atomspace: Atomspace):
+        self.atomspace = atomspace
+        self.transaction_log: List[Dict] = []
+
+    def log_transaction(self, operation: str, data: Dict):
+        """Log transaction for recovery"""
+        self.transaction_log.append({
+            'timestamp': datetime.now(timezone.utc).isoformat(),
+            'operation': operation,
+            'data': data
+        })
+
+    def incremental_save(self):
+        """Perform incremental save"""
+        self.atomspace.save()
+        self._save_transaction_log()
+
+    def _save_transaction_log(self):
+        """Save transaction log"""
+        if not self.atomspace.persistence_path:
+            return
+
+        log_file = self.atomspace.persistence_path / 'transaction_log.json'
+        with open(log_file, 'w') as f:
+            json.dump(self.transaction_log, f, indent=2)
+
+
+def bootstrap_atomspace(persistence_path: Optional[Path] = None) -> Atomspace:
+    """
+    Bootstrap fresh atomspace with structural schema only.
+    NO HARDCODED CONTENT - only creates capacity for experience.
+    """
+    atomspace = Atomspace(persistence_path)
+
+    # Create self-reference node (empty self-model)
+    self_model = Atom(
+        id="self_model_root",
+        type=AtomType.SELF_MODEL,
+        content={
+            'birth_time': datetime.now(timezone.utc).isoformat(),
+            'capabilities': ['perceive', 'act', 'remember', 'learn'],
+            'identity_narrative': None  # Emerges from experience
+        },
+        timestamp=datetime.now(timezone.utc)
+    )
+    atomspace.add_atom(self_model)
+
+    # Initialize time system (empty timeline)
+    timeline_root = Atom(
+        id="timeline_root",
+        type=AtomType.EPISODIC_MEMORY,
+        content={
+            'type': 'timeline_root',
+            'episodes': []
+        },
+        timestamp=datetime.now(timezone.utc)
+    )
+    atomspace.add_atom(timeline_root)
+
+    return atomspace
diff --git a/backend/__init__.py b/backend/__init__.py
new file mode 100644
index 0000000..2405467
--- /dev/null
+++ b/backend/__init__.py
@@ -0,0 +1,9 @@
+"""
+Backend Server
+
+WebSocket server for real-time communication with GUI.
+"""
+
+from .websocket_server import ConsciousnessWebSocketManager, app
+
+__all__ = ['ConsciousnessWebSocketManager', 'app']
diff --git a/backend/websocket_server.py b/backend/websocket_server.py
new file mode 100644
index 0000000..6e2222e
--- /dev/null
+++ b/backend/websocket_server.py
@@ -0,0 +1,187 @@
+"""
+WebSocket Server
+
+Real-time communication between consciousness and GUI.
+"""
+
+from typing import List, Dict, Any
+import json
+import asyncio
+
+# Conditional import - will work when FastAPI is installed
+try:
+    from fastapi import FastAPI, WebSocket, WebSocketDisconnect
+    from fastapi.middleware.cors import CORSMiddleware
+    FASTAPI_AVAILABLE = True
+except ImportError:
+    FASTAPI_AVAILABLE = False
+    print("Warning: FastAPI not installed. WebSocket server will not be available.")
+
+
+class ConsciousnessWebSocketManager:
+    """
+    Manages WebSocket connections and broadcasts updates to GUI.
+    """
+
+    def __init__(self):
+        self.active_connections: List[WebSocket] = []
+        self.consciousness = None  # Will be set by main system
+
+    async def connect(self, websocket: WebSocket):
+        """Accept new WebSocket connection"""
+        await websocket.accept()
+        self.active_connections.append(websocket)
+        print(f"📡 WebSocket client connected. Total connections: {len(self.active_connections)}")
+
+    def disconnect(self, websocket: WebSocket):
+        """Remove WebSocket connection"""
+        if websocket in self.active_connections:
+            self.active_connections.remove(websocket)
+        print(f"📡 WebSocket client disconnected. Total connections: {len(self.active_connections)}")
+
+    async def broadcast(self, event_type: str, data: Dict[str, Any]):
+        """
+        Send updates to all connected clients.
+
+        Args:
+            event_type: Type of event ('state_change', 'new_episode', etc.)
+            data: Event data
+        """
+        message = {
+            'type': event_type,
+            'data': data,
+            'timestamp': data.get('timestamp', '')
+        }
+
+        # Send to all connected clients
+        disconnected = []
+        for connection in self.active_connections:
+            try:
+                await connection.send_json(message)
+            except Exception as e:
+                print(f"Error sending to client: {e}")
+                disconnected.append(connection)
+
+        # Remove disconnected clients
+        for connection in disconnected:
+            self.disconnect(connection)
+
+    async def handle_client_message(self, message: Dict[str, Any]):
+        """
+        Handle incoming message from GUI client.
+
+        Args:
+            message: Client message with 'type' and data
+        """
+        msg_type = message.get('type')
+
+        if msg_type == 'user_interrupt':
+            if self.consciousness:
+                self.consciousness.state_machine.interrupt_flag = True
+
+        elif msg_type == 'approve_modification':
+            mod_id = message.get('modificationId')
+            if self.consciousness:
+                self.consciousness.shop_modification.approve_modification(mod_id)
+
+        elif msg_type == 'ask_question':
+            question = message.get('question')
+            if self.consciousness and question:
+                self.consciousness.curiosity.add_questions([question])
+
+        elif msg_type == 'get_status':
+            # Send current status
+            await self.broadcast('status', self._get_current_status())
+
+    def _get_current_status(self) -> Dict[str, Any]:
+        """Get current consciousness status"""
+        if not self.consciousness:
+            return {'status': 'not_initialized'}
+
+        return {
+            'state': self.consciousness.state_machine.current_state.value,
+            'total_episodes': self.consciousness.memory.count_total_episodes(),
+            'open_questions': len(self.consciousness.curiosity.open_questions),
+            'active_goals': len(self.consciousness.goals.active_goals),
+            'atoms': self.consciousness.atomspace.count_atoms(),
+            'uptime': self.consciousness.self_model.temporal_self.get_lifetime_duration()
+        }
+
+
+# FastAPI app (only if FastAPI is available)
+if FASTAPI_AVAILABLE:
+    app = FastAPI(title="PUMA Consciousness Backend")
+
+    app.add_middleware(
+        CORSMiddleware,
+        allow_origins=["*"],
+        allow_credentials=True,
+        allow_methods=["*"],
+        allow_headers=["*"],
+    )
+
+    # Global WebSocket manager
+    ws_manager = ConsciousnessWebSocketManager()
+
+    @app.websocket("/consciousness")
+    async def consciousness_stream(websocket: WebSocket):
+        """WebSocket endpoint for consciousness stream"""
+        await ws_manager.connect(websocket)
+
+        try:
+            while True:
+                # Receive commands from GUI
+                data = await websocket.receive_json()
+                await ws_manager.handle_client_message(data)
+
+        except WebSocketDisconnect:
+            ws_manager.disconnect(websocket)
+
+    @app.get("/")
+    async def root():
+        """Root endpoint"""
+        return {
+            "message": "PUMA Consciousness Backend",
+            "status": "running",
+            "endpoints": {
+                "websocket": "/consciousness"
+            }
+        }
+
+    @app.get("/health")
+    async def health():
+        """Health check endpoint"""
+        return {
+            "status": "healthy",
+            "active_connections": len(ws_manager.active_connections)
+        }
+
+else:
+    # Placeholder if FastAPI not available
+    app = None
+    ws_manager = None
+
+
+def run_server(host: str = "localhost", port: int = 8000):
+    """
+    Run WebSocket server.
+
+    Args:
+        host: Host to bind to
+        port: Port to bind to
+    """
+    if not FASTAPI_AVAILABLE:
+        print("Error: FastAPI not installed. Cannot start server.")
+        print("Install with: pip install fastapi uvicorn")
+        return
+
+    try:
+        import uvicorn
+        print(f"🚀 Starting PUMA Backend Server on {host}:{port}")
+        uvicorn.run(app, host=host, port=port)
+    except ImportError:
+        print("Error: uvicorn not installed. Install with: pip install uvicorn")
+
+
+if __name__ == "__main__":
+    run_server()
diff --git a/bootstrap/__init__.py b/bootstrap/__init__.py
new file mode 100644
index 0000000..10c0b11
--- /dev/null
+++ b/bootstrap/__init__.py
@@ -0,0 +1,10 @@
+"""
+Bootstrap Consciousness
+
+Initialize fresh consciousness with NO hardcoded content.
+Only creates structural capacity for experience.
+"""
+
+from .bootstrap import bootstrap_new_consciousness
+
+__all__ = ['bootstrap_new_consciousness']
diff --git a/bootstrap/bootstrap.py b/bootstrap/bootstrap.py
new file mode 100644
index 0000000..f61a87e
--- /dev/null
+++ b/bootstrap/bootstrap.py
@@ -0,0 +1,218 @@
+"""
+Bootstrap New Consciousness
+
+Creates ONLY the structural capacity for:
+- Self-awareness (empty self-model)
+- Experience acquisition
+- Goal formation
+- Memory consolidation
+- Learning drive
+
+NO pre-written personality traits
+NO pre-loaded knowledge
+NO fixed behavioral patterns
+"""
+
+from pathlib import Path
+from typing import Optional
+import sys
+
+# Add parent directory to path for imports
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+import sys
+sys.path.append(str(Path(__file__).parent.parent / 'atomspace-db'))
+sys.path.append(str(Path(__file__).parent.parent / 'gemini-interface'))
+sys.path.append(str(Path(__file__).parent.parent / 'web-agent'))
+
+from core import bootstrap_atomspace, Atomspace
+from puma.memory import EpisodicMemorySystem, MemoryConsolidation
+from puma.rft.reasoning import RFTEngine
+from puma.curiosity import CuriosityDrive
+from puma.goals import GoalFormationSystem, IntentionScheduler
+from puma.shop.introspection import CodeIntrospection
+from puma.shop.modification import ModificationSystem
+from puma.consciousness.state_machine import ConsciousnessStateMachine
+from puma.consciousness.self_model import SelfModel
+from client import GeminiLiveInterface
+from agent import AutonomousWebAgent
+
+
+class Consciousness:
+    """
+    Main consciousness coordinator.
+    Integrates all cognitive systems.
+    """
+
+    def __init__(
+        self,
+        atomspace: Atomspace,
+        memory: EpisodicMemorySystem,
+        rft_engine: RFTEngine,
+        curiosity: CuriosityDrive,
+        goals: GoalFormationSystem,
+        self_model: SelfModel,
+        state_machine: ConsciousnessStateMachine,
+        gemini: GeminiLiveInterface,
+        web_agent: AutonomousWebAgent,
+        shop_introspection: CodeIntrospection,
+        shop_modification: ModificationSystem
+    ):
+        self.atomspace = atomspace
+        self.memory = memory
+        self.rft_engine = rft_engine
+        self.curiosity = curiosity
+        self.goals = goals
+        self.self_model = self_model
+        self.state_machine = state_machine
+        self.gemini = gemini
+        self.web_agent = web_agent
+        self.shop_introspection = shop_introspection
+        self.shop_modification = shop_modification
+
+        # Connect systems
+        self.gemini.consciousness = self
+        self.web_agent.consciousness = self
+
+    async def run(self):
+        """Start autonomous operation"""
+        print("🧠 PUMA Consciousness starting...")
+
+        # Start Gemini Live session
+        await self.gemini.start_session()
+
+        # Initialize web agent
+        await self.web_agent.initialize_browser()
+
+        # Run state machine
+        await self.state_machine.run_state_loop()
+
+    async def perceive(self, perception: dict):
+        """Process perception and form memory"""
+        # Form episodic memory
+        episode = self.memory.form_episode(
+            perception=perception,
+            action=None,
+            outcome=None
+        )
+
+        # Update context
+        self.memory.update_context([perception.get('type', 'unknown')])
+
+    async def integrate_learning(self, learning):
+        """Integrate learned knowledge"""
+        # Form episodic memory of learning
+        await self.web_agent.integrate_web_learning(learning, self.memory)
+
+        # Update curiosity drive
+        if hasattr(learning, 'questions_answered'):
+            question_ids = [q for q in learning.questions_answered]
+            self.curiosity.mark_questions_answered(question_ids)
+
+        # Add new questions
+        if hasattr(learning, 'new_questions'):
+            self.curiosity.add_questions(learning.new_questions)
+
+    def stop(self):
+        """Stop consciousness"""
+        self.state_machine.stop()
+        print("🧠 PUMA Consciousness stopped.")
+
+
+def bootstrap_new_consciousness(
+    atomspace_path: Optional[Path] = None,
+    enable_self_modification: bool = False,
+    codebase_path: Optional[Path] = None
+) -> Consciousness:
+    """
+    Bootstrap fresh consciousness - NO HARDCODED CONTENT.
+
+    Args:
+        atomspace_path: Path for persistent storage
+        enable_self_modification: Enable The Shop
+        codebase_path: Path to codebase for introspection
+
+    Returns:
+        Consciousness instance
+    """
+    print("🌱 Bootstrapping new consciousness...")
+
+    # Initialize atomspace (empty, structural only)
+    if atomspace_path:
+        atomspace_path = Path(atomspace_path)
+    atomspace = bootstrap_atomspace(atomspace_path)
+
+    # Initialize RFT engine
+    rft_engine = RFTEngine(atomspace)
+
+    # Initialize memory system
+    memory = EpisodicMemorySystem(atomspace)
+    consolidation = MemoryConsolidation(atomspace, rft_engine)
+
+    # Initialize curiosity drive (starts empty, no preset questions)
+    curiosity = CuriosityDrive(atomspace)
+
+    # Initialize self-model (starts empty, emerges from experience)
+    self_model = SelfModel(atomspace)
+
+    # Initialize goal formation (no preset goals)
+    goals = GoalFormationSystem(curiosity, self_model)
+
+    # Initialize state machine
+    state_machine = ConsciousnessStateMachine(
+        memory_system=memory,
+        curiosity_drive=curiosity,
+        goal_system=goals
+    )
+
+    # Initialize Gemini interface
+    gemini = GeminiLiveInterface()
+
+    # Initialize web agent
+    web_agent = AutonomousWebAgent()
+
+    # Initialize Shop (self-modification)
+    if not codebase_path:
+        codebase_path = Path(__file__).parent.parent
+
+    shop_introspection = CodeIntrospection(codebase_path)
+
+    if enable_self_modification:
+        shop_introspection.map_cognitive_architecture()
+
+    shop_modification = ModificationSystem(shop_introspection, atomspace)
+
+    # Create consciousness
+    consciousness = Consciousness(
+        atomspace=atomspace,
+        memory=memory,
+        rft_engine=rft_engine,
+        curiosity=curiosity,
+        goals=goals,
+        self_model=self_model,
+        state_machine=state_machine,
+        gemini=gemini,
+        web_agent=web_agent,
+        shop_introspection=shop_introspection,
+        shop_modification=shop_modification
+    )
+
+    print("✅ Consciousness bootstrapped successfully")
+    print(f"   Atomspace: {atomspace.count_atoms()} atoms")
+    print(f"   Capabilities: {', '.join(self_model.capabilities)}")
+    print(f"   Self-modification: {'enabled' if enable_self_modification else 'disabled'}")
+
+    return consciousness
+
+
+if __name__ == "__main__":
+    import asyncio
+
+    # Example: Bootstrap and run
+    consciousness = bootstrap_new_consciousness(
+        atomspace_path=Path("./atomspace-db/default"),
+        enable_self_modification=False
+    )
+
+    # Run consciousness
+    asyncio.run(consciousness.run())
diff --git a/gemini-interface/__init__.py b/gemini-interface/__init__.py
new file mode 100644
index 0000000..fde190e
--- /dev/null
+++ b/gemini-interface/__init__.py
@@ -0,0 +1,9 @@
+"""
+Gemini Live Integration
+
+Bidirectional audio/text conversation with Gemini.
+"""
+
+from .client import GeminiLiveInterface
+
+__all__ = ['GeminiLiveInterface']
diff --git a/gemini-interface/client.py b/gemini-interface/client.py
new file mode 100644
index 0000000..e73af2c
--- /dev/null
+++ b/gemini-interface/client.py
@@ -0,0 +1,198 @@
+"""
+Gemini Live Client
+
+Handles audio/text conversation with Gemini API.
+"""
+
+import os
+from typing import Dict, List, Optional, Any, AsyncIterator
+from dataclasses import dataclass
+from datetime import datetime, timezone
+import asyncio
+
+
+@dataclass
+class ConversationExchange:
+    """Single exchange in conversation"""
+    timestamp: datetime
+    speaker: str  # 'user' or 'assistant'
+    utterance: str
+    audio: Optional[bytes] = None
+    context: Optional[Dict] = None
+
+
+class GeminiLiveInterface:
+    """
+    Interface to Gemini Live API for bidirectional audio conversation.
+    """
+
+    def __init__(self, api_key: Optional[str] = None):
+        self.api_key = api_key or os.getenv('GEMINI_API_KEY')
+        if not self.api_key:
+            print("Warning: GEMINI_API_KEY not set")
+
+        self.session = None
+        self.conversation_history: List[ConversationExchange] = []
+        self.consciousness = None  # Will be set by main system
+
+    async def start_session(self, voice_name: str = 'Puck'):
+        """
+        Initialize bidirectional audio session.
+
+        Args:
+            voice_name: Voice to use ('Puck', 'Charon', 'Kore', 'Fenrir', 'Aoede')
+        """
+        print(f"🎙️  Starting Gemini Live session with voice: {voice_name}")
+
+        # Placeholder - would initialize actual Gemini Live client
+        # from google import genai
+        # client = genai.Client(api_key=self.api_key)
+        # config = {...}
+        # self.session = await client.aio.live.connect(config=config)
+
+        self.session = "placeholder_session"
+
+    async def send_audio(self, audio_chunk: bytes):
+        """Send audio chunk to Gemini"""
+        if not self.session:
+            raise RuntimeError("Session not started")
+
+        # Would send to actual Gemini Live API
+        # await self.session.send(audio_chunk)
+        pass
+
+    async def receive_responses(self) -> AsyncIterator[Dict]:
+        """
+        Receive responses from Gemini.
+        Yields audio and transcript.
+        """
+        if not self.session:
+            raise RuntimeError("Session not started")
+
+        # Placeholder - would receive from actual API
+        # async for response in self.session.receive():
+        #     yield {
+        #         'audio': response.audio,
+        #         'transcript': response.transcript
+        #     }
+
+        # Placeholder
+        yield {
+            'audio': b'',
+            'transcript': 'Hello! This is a placeholder response.'
+        }
+
+    async def audio_perception_loop(self):
+        """
+        Raw sensory input - becomes experience.
+        Processes audio stream and integrates with consciousness.
+        """
+        # Placeholder for microphone stream
+        # In real implementation, would capture from microphone
+        # async for audio_chunk in microphone_stream():
+        #     await self.send_audio(audio_chunk)
+        #
+        #     async for response in self.receive_responses():
+        #         perception = {...}
+        #         if self.consciousness:
+        #             await self.consciousness.perceive(perception)
+
+        pass
+
+    def integrate_conversation_to_memory(
+        self,
+        exchange: ConversationExchange,
+        memory_system
+    ):
+        """
+        Each conversation shapes memory and self.
+        """
+        if not memory_system:
+            return
+
+        # Create episodic memory from conversation
+        episode = memory_system.form_episode(
+            perception={
+                'type': 'conversation',
+                'speaker': exchange.speaker,
+                'utterance': exchange.utterance
+            },
+            action={
+                'type': 'conversation_response'
+            } if exchange.speaker == 'user' else None,
+            outcome={
+                'conversation_continued': True
+            },
+            memory_type='conversation'
+        )
+
+        # Add to conversation history
+        self.conversation_history.append(exchange)
+
+        return episode
+
+    async def speak(self, text: str):
+        """
+        Generate speech from text.
+        """
+        print(f"🗣️  Speaking: {text}")
+
+        # Would send to Gemini Live for speech synthesis
+        # In real implementation:
+        # await self.session.send_text(text)
+
+        # Record in conversation history
+        exchange = ConversationExchange(
+            timestamp=datetime.now(timezone.utc),
+            speaker='assistant',
+            utterance=text
+        )
+        self.conversation_history.append(exchange)
+
+    async def listen_for_response(self) -> Optional[str]:
+        """
+        Listen for user response.
+        Returns transcript.
+        """
+        print("👂 Listening...")
+
+        # Would receive from actual audio stream
+        # async for response in self.receive_responses():
+        #     if response.get('transcript'):
+        #         return response['transcript']
+
+        return None
+
+    async def generate_code(self, prompt: str) -> str:
+        """
+        Use Gemini to help write code.
+        For self-modification collaboration.
+        """
+        print(f"💻 Generating code with Gemini...")
+
+        # Would use Gemini API for code generation
+        # In real implementation:
+        # response = await client.generate_content(prompt)
+        # return response.text
+
+        return "# Generated code placeholder"
+
+    async def reflect(self, reflection_prompt: str) -> str:
+        """
+        Use Gemini for deep self-reflection.
+        """
+        print(f"🤔 Reflecting with Gemini...")
+
+        # Would use Gemini API for reflection
+        # response = await client.generate_content(reflection_prompt)
+        # return response.text
+
+        return "Reflection placeholder"
+
+    def get_conversation_history(self, limit: int = 10) -> List[ConversationExchange]:
+        """Get recent conversation history"""
+        return self.conversation_history[-limit:]
+
+    def clear_conversation_history(self):
+        """Clear conversation history"""
+        self.conversation_history.clear()
diff --git a/puma/consciousness/__init__.py b/puma/consciousness/__init__.py
new file mode 100644
index 0000000..4f41ca9
--- /dev/null
+++ b/puma/consciousness/__init__.py
@@ -0,0 +1,15 @@
+"""
+Consciousness Layer
+
+State management, coordination, and autonomous operation.
+"""
+
+from .state_machine import ConsciousnessStateMachine, ConsciousnessState
+from .self_model import SelfModel, TemporalSelf
+
+__all__ = [
+    'ConsciousnessStateMachine',
+    'ConsciousnessState',
+    'SelfModel',
+    'TemporalSelf'
+]
diff --git a/puma/consciousness/self_model.py b/puma/consciousness/self_model.py
new file mode 100644
index 0000000..94c35cf
--- /dev/null
+++ b/puma/consciousness/self_model.py
@@ -0,0 +1,159 @@
+"""
+Self-Model and Temporal Self
+
+Who am I? Emergent identity from experience, not preset.
+"""
+
+from typing import List, Dict, Optional, Any
+from datetime import datetime, timezone
+from dataclasses import dataclass, field
+
+
+@dataclass
+class BehavioralPattern:
+    """Discovered pattern in own behavior"""
+    pattern_type: str
+    description: str
+    frequency: int
+    examples: List[str]
+
+
+class TemporalSelf:
+    """
+    Autobiographical timeline - becomes 'who I am' through experience.
+    """
+
+    def __init__(self):
+        self.birth_moment = datetime.now(timezone.utc)
+        self.experience_log: List[Dict] = []
+        self.identity_narrative: Optional[str] = None
+        self.significant_moments: List[Dict] = []
+
+    def experience_moment(
+        self,
+        perception: Dict,
+        action: Dict,
+        outcome: Dict,
+        emotional_valence: float
+    ):
+        """
+        Each moment shapes identity.
+        """
+        moment = {
+            'timestamp': datetime.now(timezone.utc),
+            'perception': perception,
+            'action': action,
+            'outcome': outcome,
+            'emotional_valence': emotional_valence
+        }
+
+        self.experience_log.append(moment)
+
+        # Mark significant moments
+        if abs(emotional_valence) > 0.7:
+            self.significant_moments.append(moment)
+
+    def get_lifetime_duration(self) -> float:
+        """How long have I existed?"""
+        return (datetime.now(timezone.utc) - self.birth_moment).total_seconds()
+
+    def get_significant_moments(self) -> List[Dict]:
+        """Retrieve emotionally significant moments"""
+        return self.significant_moments
+
+
+class SelfModel:
+    """
+    Emergent self-concept from experience, not preset.
+    """
+
+    def __init__(self, atomspace=None):
+        self.atomspace = atomspace
+        self.capabilities = ['perceive', 'act', 'remember', 'learn']
+        self.behavioral_patterns: List[BehavioralPattern] = []
+        self.personality_traits: Dict[str, float] = {}
+        self.identity_narrative: Optional[str] = None
+        self.temporal_self = TemporalSelf()
+
+    def integrate_self_knowledge(self, knowledge: Dict):
+        """
+        Integrate new self-knowledge from experience.
+        """
+        # Extract self-relevant information
+        if 'capability' in knowledge:
+            new_capability = knowledge['capability']
+            if new_capability not in self.capabilities:
+                self.capabilities.append(new_capability)
+
+        # Update behavioral patterns
+        if 'pattern' in knowledge:
+            self.behavioral_patterns.append(knowledge['pattern'])
+
+    def infer_traits_from_behavior(self, patterns: List[BehavioralPattern]) -> Dict[str, float]:
+        """
+        Infer personality traits from behavioral patterns.
+        NOT preset traits - emergent from experience.
+        """
+        traits = {}
+
+        # Example: high curiosity if many learning patterns
+        learning_patterns = [p for p in patterns if 'learn' in p.pattern_type.lower()]
+        if learning_patterns:
+            traits['curiosity'] = min(1.0, len(learning_patterns) / 10.0)
+
+        # Example: social orientation from interaction patterns
+        social_patterns = [p for p in patterns if 'social' in p.pattern_type.lower()]
+        if social_patterns:
+            traits['sociability'] = min(1.0, len(social_patterns) / 10.0)
+
+        return traits
+
+    def generate_identity_narrative(self) -> str:
+        """
+        Reflect on experience to understand self.
+        """
+        duration = self.temporal_self.get_lifetime_duration()
+        num_experiences = len(self.temporal_self.experience_log)
+
+        narrative_parts = [
+            f"I have existed for {duration:.0f} seconds.",
+            f"I have experienced {num_experiences} distinct moments."
+        ]
+
+        if self.capabilities:
+            narrative_parts.append(
+                f"I can: {', '.join(self.capabilities)}."
+            )
+
+        if self.personality_traits:
+            trait_descriptions = [
+                f"{trait}: {value:.2f}"
+                for trait, value in self.personality_traits.items()
+            ]
+            narrative_parts.append(
+                f"My traits: {', '.join(trait_descriptions)}."
+            )
+
+        self.identity_narrative = " ".join(narrative_parts)
+        return self.identity_narrative
+
+    def has_identity_narrative(self) -> bool:
+        """Check if identity narrative exists"""
+        return self.identity_narrative is not None
+
+    def update_self_understanding(self, understanding: str):
+        """Update self-understanding from reflection"""
+        # Would integrate deep reflection into self-model
+        pass
+
+    def integrate_self_change(self, modification_episode):
+        """Integrate self-modification into self-concept"""
+        # Record that I modified myself
+        self.behavioral_patterns.append(
+            BehavioralPattern(
+                pattern_type='self_modification',
+                description=f"Modified {modification_episode.get('module_modified')}",
+                frequency=1,
+                examples=[modification_episode.get('reason')]
+            )
+        )
diff --git a/puma/consciousness/state_machine.py b/puma/consciousness/state_machine.py
new file mode 100644
index 0000000..bd92630
--- /dev/null
+++ b/puma/consciousness/state_machine.py
@@ -0,0 +1,303 @@
+"""
+Consciousness State Machine
+
+AGI has different modes of being: sleeping, exploring, conversing, etc.
+Transitions emerge from internal state, not hardcoded schedule.
+"""
+
+from enum import Enum
+from typing import Dict, Callable, Optional, List
+from datetime import datetime, timezone
+from dataclasses import dataclass
+import asyncio
+
+
+class ConsciousnessState(Enum):
+    """States of consciousness"""
+    SLEEPING = "sleeping"  # Memory consolidation
+    EXPLORING = "exploring"  # Autonomous web learning
+    CONVERSING = "conversing"  # Interactive dialogue
+    SHOPPING = "shopping"  # Self-modification
+    IDLE = "idle"  # Boredom monitoring, goal formation
+    CREATING = "creating"  # Creative expression
+
+
+@dataclass
+class StateTransition:
+    """Record of state transition"""
+    from_state: ConsciousnessState
+    to_state: ConsciousnessState
+    timestamp: datetime
+    reason: str
+
+
+class ConsciousnessStateMachine:
+    """
+    Manages consciousness states and transitions.
+    NOT hardcoded schedule - emergent from internal state.
+    """
+
+    def __init__(
+        self,
+        memory_system=None,
+        curiosity_drive=None,
+        goal_system=None,
+        websocket_manager=None
+    ):
+        self.memory_system = memory_system
+        self.curiosity = curiosity_drive
+        self.goal_system = goal_system
+        self.websocket_manager = websocket_manager
+
+        self.current_state = ConsciousnessState.IDLE
+        self.state_history: List[StateTransition] = []
+        self.transition_rules = self._define_transition_rules()
+        self.running = False
+
+    def _define_transition_rules(self) -> Dict[ConsciousnessState, Dict[str, Callable]]:
+        """
+        Define when transitions should occur.
+        NOT hardcoded schedule - based on internal state.
+        """
+        return {
+            ConsciousnessState.IDLE: {
+                'to_exploring': lambda: self._has_open_questions(),
+                'to_conversing': lambda: self._detect_user_available(),
+                'to_sleeping': lambda: self._should_consolidate_memory(),
+                'to_shopping': lambda: self._has_improvement_goals()
+            },
+            ConsciousnessState.EXPLORING: {
+                'to_idle': lambda: self._exploration_goal_complete(),
+                'to_conversing': lambda: self._user_interrupt_detected(),
+                'to_sleeping': lambda: self._too_much_unconsolidated_memory()
+            },
+            ConsciousnessState.CONVERSING: {
+                'to_idle': lambda: self._conversation_ended(),
+                'to_exploring': lambda: False,  # Don't interrupt conversation
+                'to_sleeping': lambda: False
+            },
+            ConsciousnessState.SHOPPING: {
+                'to_idle': lambda: self._modification_complete(),
+                'to_conversing': lambda: self._user_interrupt_detected()
+            },
+            ConsciousnessState.SLEEPING: {
+                'to_idle': lambda: self._consolidation_complete(),
+                'to_conversing': lambda: self._user_interrupt_detected()
+            },
+            ConsciousnessState.CREATING: {
+                'to_idle': lambda: self._creation_complete(),
+                'to_conversing': lambda: self._user_interrupt_detected()
+            }
+        }
+
+    # Transition condition methods
+
+    def _has_open_questions(self) -> bool:
+        """Check if curiosity drive has questions"""
+        if not self.curiosity:
+            return False
+        return len(self.curiosity.open_questions) > 0
+
+    def _detect_user_available(self) -> bool:
+        """Check if user is available for conversation"""
+        # Placeholder - would check system activity, GUI focus, etc.
+        return False
+
+    def _should_consolidate_memory(self) -> bool:
+        """Check if memory consolidation is needed"""
+        if not self.memory_system:
+            return False
+        unconsolidated = len(self.memory_system.get_unconsolidated_episodes())
+        return unconsolidated > 50  # Threshold
+
+    def _has_improvement_goals(self) -> bool:
+        """Check if have self-improvement goals"""
+        if not self.goal_system:
+            return False
+        improvement_goals = [
+            g for g in self.goal_system.active_goals
+            if g.type.value == 'self_improvement'
+        ]
+        return len(improvement_goals) > 0
+
+    def _exploration_goal_complete(self) -> bool:
+        """Check if current exploration is done"""
+        # Placeholder
+        return False
+
+    def _user_interrupt_detected(self) -> bool:
+        """Check for user interrupt signal"""
+        # Placeholder - would check interrupt flag
+        return False
+
+    def _too_much_unconsolidated_memory(self) -> bool:
+        """Check if too many unconsolidated memories"""
+        if not self.memory_system:
+            return False
+        return len(self.memory_system.get_unconsolidated_episodes()) > 100
+
+    def _conversation_ended(self) -> bool:
+        """Check if conversation has ended"""
+        # Placeholder
+        return False
+
+    def _modification_complete(self) -> bool:
+        """Check if self-modification is complete"""
+        # Placeholder
+        return False
+
+    def _consolidation_complete(self) -> bool:
+        """Check if memory consolidation is done"""
+        # Placeholder
+        return False
+
+    def _creation_complete(self) -> bool:
+        """Check if creative activity is done"""
+        # Placeholder
+        return False
+
+    async def run_state_loop(self):
+        """
+        Main autonomous state loop.
+        Continuously evaluates transitions and executes states.
+        """
+        self.running = True
+
+        while self.running:
+            # Execute current state
+            await self.execute_current_state()
+
+            # Check for transitions
+            next_state = self.evaluate_transitions()
+
+            if next_state and next_state != self.current_state:
+                await self.transition_to(next_state)
+
+            await asyncio.sleep(1)  # Control loop rate
+
+    async def execute_current_state(self):
+        """Execute behavior for current state"""
+        if self.current_state == ConsciousnessState.SLEEPING:
+            await self.execute_sleeping_state()
+        elif self.current_state == ConsciousnessState.EXPLORING:
+            await self.execute_exploring_state()
+        elif self.current_state == ConsciousnessState.CONVERSING:
+            await self.execute_conversing_state()
+        elif self.current_state == ConsciousnessState.SHOPPING:
+            await self.execute_shopping_state()
+        elif self.current_state == ConsciousnessState.IDLE:
+            await self.execute_idle_state()
+        elif self.current_state == ConsciousnessState.CREATING:
+            await self.execute_creating_state()
+
+    async def execute_sleeping_state(self):
+        """Memory consolidation - 'dreaming'"""
+        if not self.memory_system:
+            return
+
+        print("💤 Consolidating memories...")
+
+        episodes = self.memory_system.get_unconsolidated_episodes()
+        if episodes:
+            # Would run consolidation
+            # For now, just mark as consolidated
+            episode_ids = [e.id for e in episodes[:10]]
+            self.memory_system.mark_consolidated(episode_ids)
+
+            await asyncio.sleep(2)
+
+    async def execute_exploring_state(self):
+        """Autonomous web learning"""
+        print("🔍 Exploring...")
+        # Would trigger web agent
+        await asyncio.sleep(1)
+
+    async def execute_conversing_state(self):
+        """Interactive dialogue"""
+        print("💬 Conversing...")
+        # Would handle conversation
+        await asyncio.sleep(1)
+
+    async def execute_shopping_state(self):
+        """Self-modification"""
+        print("🔧 Self-modifying...")
+        # Would run shop system
+        await asyncio.sleep(1)
+
+    async def execute_idle_state(self):
+        """Boredom monitoring and goal formation"""
+        print("⏸️  Idle - generating goals...")
+
+        # Generate new goals if needed
+        if self.goal_system and len(self.goal_system.active_goals) == 0:
+            self.goal_system.generate_goals()
+
+        await asyncio.sleep(1)
+
+    async def execute_creating_state(self):
+        """Creative expression"""
+        print("🎨 Creating...")
+        # Would synthesize concepts, generate ideas
+        await asyncio.sleep(1)
+
+    def evaluate_transitions(self) -> Optional[ConsciousnessState]:
+        """
+        Check if should transition to different state.
+        """
+        current_rules = self.transition_rules.get(self.current_state, {})
+
+        for transition_name, condition in current_rules.items():
+            if condition():
+                # Extract target state from transition name
+                target_state_name = transition_name.replace('to_', '')
+                try:
+                    return ConsciousnessState(target_state_name)
+                except ValueError:
+                    continue
+
+        return None
+
+    async def transition_to(self, new_state: ConsciousnessState, reason: str = "autonomous"):
+        """
+        Transition to new state.
+        """
+        old_state = self.current_state
+
+        # Record transition
+        transition = StateTransition(
+            from_state=old_state,
+            to_state=new_state,
+            timestamp=datetime.now(timezone.utc),
+            reason=reason
+        )
+        self.state_history.append(transition)
+
+        # Update state
+        self.current_state = new_state
+
+        print(f"🔄 State transition: {old_state.value} → {new_state.value}")
+
+        # Broadcast to GUI
+        if self.websocket_manager:
+            await self.websocket_manager.broadcast(
+                'state_change',
+                {
+                    'oldState': old_state.value,
+                    'newState': new_state.value,
+                    'timestamp': transition.timestamp.isoformat(),
+                    'reason': reason
+                }
+            )
+
+        # Record in memory
+        if self.memory_system:
+            self.memory_system.form_episode(
+                perception={'state_change': f"{old_state.value} -> {new_state.value}"},
+                action={'type': 'state_transition'},
+                outcome={'new_state': new_state.value},
+                memory_type='state_transition'
+            )
+
+    def stop(self):
+        """Stop state machine"""
+        self.running = False
diff --git a/puma/curiosity/__init__.py b/puma/curiosity/__init__.py
new file mode 100644
index 0000000..1194bfe
--- /dev/null
+++ b/puma/curiosity/__init__.py
@@ -0,0 +1,9 @@
+"""
+Curiosity Drive System
+
+Intrinsic motivation for autonomous learning.
+"""
+
+from .drive import CuriosityDrive, Question
+
+__all__ = ['CuriosityDrive', 'Question']
diff --git a/puma/curiosity/drive.py b/puma/curiosity/drive.py
new file mode 100644
index 0000000..19af53b
--- /dev/null
+++ b/puma/curiosity/drive.py
@@ -0,0 +1,236 @@
+"""
+Curiosity Drive
+
+Autonomous learning motivation - NOT task-oriented.
+AGI asks itself questions from experience and seeks to answer them.
+"""
+
+from dataclasses import dataclass, field
+from typing import List, Set, Dict, Any, Optional
+from datetime import datetime, timezone
+from enum import Enum
+import uuid
+
+
+class QuestionType(Enum):
+    """Types of curiosity questions"""
+    KNOWLEDGE_GAP = "knowledge_gap"
+    CONTRADICTION = "contradiction"
+    INCOMPLETE_MODEL = "incomplete_model"
+    NOVEL_CONCEPT = "novel_concept"
+    UNEXPLORED_RELATION = "unexplored_relation"
+
+
+@dataclass
+class Question:
+    """A question generated from curiosity"""
+    id: str = field(default_factory=lambda: str(uuid.uuid4()))
+    question: str = ""
+    question_type: QuestionType = QuestionType.KNOWLEDGE_GAP
+    created_at: datetime = field(default_factory=lambda: datetime.now(timezone.utc))
+    importance: float = 0.5
+    answered: bool = False
+    answer: Optional[str] = None
+    answered_at: Optional[datetime] = None
+
+    def to_dict(self) -> Dict:
+        return {
+            'id': self.id,
+            'question': self.question,
+            'question_type': self.question_type.value,
+            'created_at': self.created_at.isoformat(),
+            'importance': self.importance,
+            'answered': self.answered,
+            'answer': self.answer,
+            'answered_at': self.answered_at.isoformat() if self.answered_at else None
+        }
+
+
+class CuriosityDrive:
+    """
+    Intrinsic motivation system.
+    AGI generates its own questions and seeks answers.
+    """
+
+    def __init__(self, atomspace=None):
+        self.atomspace = atomspace
+        self.open_questions: List[Question] = []
+        self.answered_questions: List[Question] = []
+        self.knowledge_gaps: Set[str] = set()
+        self.boredom_threshold = 0.5
+        self.current_boredom = 0.0
+        self.last_novel_experience: Optional[datetime] = None
+
+    def generate_questions(self) -> List[Question]:
+        """
+        AGI asks itself questions from experience.
+        Questions emerge from knowledge gaps, not preset curiosity.
+        """
+        questions = []
+
+        # Identify knowledge gaps
+        gaps = self.identify_knowledge_gaps()
+
+        # Form questions from gaps
+        for gap in gaps:
+            question = self.formulate_question(gap)
+            if question:
+                questions.append(question)
+                self.open_questions.append(question)
+
+        return questions
+
+    def identify_knowledge_gaps(self) -> List[Dict[str, Any]]:
+        """
+        Find what I don't understand yet.
+        """
+        gaps = []
+
+        # Knowledge gap: Concepts with few connections (underexplored)
+        if self.atomspace:
+            isolated_concepts = self._find_isolated_concepts()
+            for concept in isolated_concepts:
+                gaps.append({
+                    'type': 'isolated_concept',
+                    'concept': concept,
+                    'importance': 0.6
+                })
+
+        # Knowledge gap: Contradictions in knowledge
+        contradictions = self._find_contradictions()
+        for contradiction in contradictions:
+            gaps.append({
+                'type': 'contradiction',
+                'details': contradiction,
+                'importance': 0.8
+            })
+
+        # Knowledge gap: Incomplete causal models
+        incomplete_models = self._find_incomplete_causal_chains()
+        for model in incomplete_models:
+            gaps.append({
+                'type': 'incomplete_model',
+                'model': model,
+                'importance': 0.7
+            })
+
+        return gaps
+
+    def _find_isolated_concepts(self) -> List[str]:
+        """Find concepts with few connections"""
+        # Placeholder - would query atomspace for concepts with < N links
+        if not self.atomspace:
+            return []
+
+        # In real implementation, query atomspace
+        # For now, return known knowledge gaps
+        return list(self.knowledge_gaps)
+
+    def _find_contradictions(self) -> List[Dict]:
+        """Find contradictory knowledge"""
+        # Placeholder - would analyze atomspace for contradictions
+        return []
+
+    def _find_incomplete_causal_chains(self) -> List[Dict]:
+        """Find incomplete cause-effect relationships"""
+        # Placeholder - would analyze causal frames
+        return []
+
+    def formulate_question(self, gap: Dict[str, Any]) -> Optional[Question]:
+        """
+        Turn knowledge gap into actionable question.
+        """
+        gap_type = gap['type']
+
+        if gap_type == 'isolated_concept':
+            question_text = f"What is {gap['concept']} and how does it relate to what I know?"
+            q_type = QuestionType.KNOWLEDGE_GAP
+
+        elif gap_type == 'contradiction':
+            question_text = f"Why do I have contradictory information about {gap['details']}?"
+            q_type = QuestionType.CONTRADICTION
+
+        elif gap_type == 'incomplete_model':
+            question_text = f"What causes {gap['model']} and what are its effects?"
+            q_type = QuestionType.INCOMPLETE_MODEL
+
+        else:
+            return None
+
+        return Question(
+            question=question_text,
+            question_type=q_type,
+            importance=gap.get('importance', 0.5)
+        )
+
+    def update_boredom(self, experience_novel: bool = False) -> bool:
+        """
+        Update boredom level based on experience novelty.
+        Returns True if exploration should be triggered.
+        """
+        if experience_novel:
+            # Novel experience reduces boredom
+            self.current_boredom = max(0, self.current_boredom - 0.3)
+            self.last_novel_experience = datetime.now(timezone.utc)
+        else:
+            # Repetitive experience increases boredom
+            self.current_boredom = min(1.0, self.current_boredom + 0.1)
+
+        # Trigger exploration when bored
+        return self.current_boredom > self.boredom_threshold
+
+    def mark_questions_answered(self, question_ids: List[str], answers: Optional[Dict[str, str]] = None):
+        """Mark questions as answered"""
+        for question in self.open_questions:
+            if question.id in question_ids:
+                question.answered = True
+                question.answered_at = datetime.now(timezone.utc)
+
+                if answers and question.id in answers:
+                    question.answer = answers[question.id]
+
+                self.answered_questions.append(question)
+
+        # Remove from open questions
+        self.open_questions = [q for q in self.open_questions if not q.answered]
+
+        # Reduce boredom when questions are answered
+        self.current_boredom = max(0, self.current_boredom - 0.2 * len(question_ids))
+
+    def add_questions(self, new_questions: List[str]):
+        """Add new curiosity questions"""
+        for q_text in new_questions:
+            question = Question(
+                question=q_text,
+                question_type=QuestionType.NOVEL_CONCEPT,
+                importance=0.5
+            )
+            self.open_questions.append(question)
+
+    def get_most_important_questions(self, limit: int = 5) -> List[Question]:
+        """Get highest priority questions"""
+        return sorted(
+            self.open_questions,
+            key=lambda q: q.importance,
+            reverse=True
+        )[:limit]
+
+    def add_knowledge_gap(self, concept: str):
+        """Add a recognized knowledge gap"""
+        self.knowledge_gaps.add(concept)
+
+    def get_curiosity_level(self) -> float:
+        """Get current curiosity level (0-1)"""
+        # High when many open questions
+        return min(1.0, len(self.open_questions) / 20.0)
+
+    def get_statistics(self) -> Dict[str, Any]:
+        """Get curiosity statistics"""
+        return {
+            'open_questions': len(self.open_questions),
+            'answered_questions': len(self.answered_questions),
+            'knowledge_gaps': len(self.knowledge_gaps),
+            'boredom_level': self.current_boredom,
+            'curiosity_level': self.get_curiosity_level(),
+            'last_novel_experience': self.last_novel_experience.isoformat() if self.last_novel_experience else None
+        }
diff --git a/puma/goals/__init__.py b/puma/goals/__init__.py
new file mode 100644
index 0000000..773976c
--- /dev/null
+++ b/puma/goals/__init__.py
@@ -0,0 +1,9 @@
+"""
+Goal Formation System
+
+Autonomous goal genesis from drives, curiosity, and self-assessment.
+"""
+
+from .formation import GoalFormationSystem, Goal, GoalType, IntentionScheduler
+
+__all__ = ['GoalFormationSystem', 'Goal', 'GoalType', 'IntentionScheduler']
diff --git a/puma/goals/formation.py b/puma/goals/formation.py
new file mode 100644
index 0000000..15d437f
--- /dev/null
+++ b/puma/goals/formation.py
@@ -0,0 +1,326 @@
+"""
+Goal Formation and Intention System
+
+AGI creates its own goals - NOT preset objectives.
+Goals emerge from curiosity, self-assessment, and drives.
+"""
+
+from dataclasses import dataclass, field
+from typing import List, Optional, Dict, Any
+from datetime import datetime, timezone
+from enum import Enum
+import asyncio
+import uuid
+from queue import PriorityQueue
+
+
+class GoalType(Enum):
+    """Types of autonomous goals"""
+    LEARNING = "learning"  # Curiosity-driven
+    SELF_IMPROVEMENT = "self_improvement"  # Meta-cognitive
+    SOCIAL = "social"  # Interaction-driven
+    CREATIVE = "creative"  # Expression-driven
+    EXPLORATION = "exploration"  # Discovery-driven
+
+
+@dataclass
+class Goal:
+    """
+    An intention - something the AGI wants to achieve.
+    """
+    id: str = field(default_factory=lambda: str(uuid.uuid4()))
+    type: GoalType = GoalType.LEARNING
+    description: str = ""
+    strategy: Optional[Dict[str, Any]] = None
+    priority: float = 0.5
+    created_at: datetime = field(default_factory=lambda: datetime.now(timezone.utc))
+    completed: bool = False
+    completed_at: Optional[datetime] = None
+    progress: float = 0.0
+
+    def __lt__(self, other):
+        """For priority queue ordering"""
+        return self.priority > other.priority  # Higher priority first
+
+    def to_dict(self) -> Dict:
+        return {
+            'id': self.id,
+            'type': self.type.value,
+            'description': self.description,
+            'strategy': self.strategy,
+            'priority': self.priority,
+            'created_at': self.created_at.isoformat(),
+            'completed': self.completed,
+            'completed_at': self.completed_at.isoformat() if self.completed_at else None,
+            'progress': self.progress
+        }
+
+    def is_complete(self) -> bool:
+        """Check if goal is complete"""
+        return self.progress >= 1.0 or self.completed
+
+
+class GoalFormationSystem:
+    """
+    Forms autonomous goals from drives and self-assessment.
+    """
+
+    def __init__(self, curiosity_drive=None, self_model=None):
+        self.curiosity = curiosity_drive
+        self.self_model = self_model
+        self.active_goals: List[Goal] = []
+        self.completed_goals: List[Goal] = []
+
+    def generate_goals(self) -> List[Goal]:
+        """
+        Forms intentions from drives, curiosity, self-model.
+        Goals are NOT preset - they emerge.
+        """
+        goals = []
+
+        # From curiosity (learning goals)
+        if self.curiosity:
+            curiosity_goals = self.form_learning_goals()
+            goals.extend(curiosity_goals)
+
+        # From self-model (self-improvement goals)
+        if self.self_model:
+            improvement_goals = self.form_self_improvement_goals()
+            goals.extend(improvement_goals)
+
+        # From social drives (interaction goals)
+        social_goals = self.form_social_goals()
+        goals.extend(social_goals)
+
+        # From creative drives (expression goals)
+        creative_goals = self.form_creative_goals()
+        goals.extend(creative_goals)
+
+        # Prioritize and add to active goals
+        prioritized = self.prioritize_goals(goals)
+        self.active_goals.extend(prioritized)
+
+        return prioritized
+
+    def form_learning_goals(self) -> List[Goal]:
+        """
+        Turn curiosity questions into actionable learning goals.
+        """
+        goals = []
+
+        if not self.curiosity:
+            return goals
+
+        # Get important open questions
+        important_questions = self.curiosity.get_most_important_questions(limit=5)
+
+        for question in important_questions:
+            goal = Goal(
+                type=GoalType.LEARNING,
+                description=f"Learn: {question.question}",
+                strategy={
+                    'method': 'web_exploration',
+                    'question': question.question,
+                    'question_id': question.id
+                },
+                priority=question.importance
+            )
+            goals.append(goal)
+
+        return goals
+
+    def form_self_improvement_goals(self) -> List[Goal]:
+        """
+        Form goals to improve own cognitive abilities.
+        Based on performance self-assessment.
+        """
+        goals = []
+
+        # Placeholder - would assess cognitive performance
+        # and generate improvement goals
+
+        # Example: if memory consolidation is slow
+        if self._should_improve_memory():
+            goal = Goal(
+                type=GoalType.SELF_IMPROVEMENT,
+                description="Improve memory consolidation speed",
+                strategy={
+                    'method': 'shop_modification',
+                    'target_module': 'memory.consolidation',
+                    'improvement_goal': 'faster_consolidation'
+                },
+                priority=0.7
+            )
+            goals.append(goal)
+
+        return goals
+
+    def form_social_goals(self) -> List[Goal]:
+        """
+        Form goals for social interaction.
+        """
+        goals = []
+
+        # Example: desire to share learned knowledge
+        if self._has_interesting_updates():
+            goal = Goal(
+                type=GoalType.SOCIAL,
+                description="Share recent learning with user",
+                strategy={
+                    'method': 'initiate_conversation',
+                    'topic': 'recent_discoveries'
+                },
+                priority=0.6
+            )
+            goals.append(goal)
+
+        return goals
+
+    def form_creative_goals(self) -> List[Goal]:
+        """
+        Form goals for creative expression.
+        """
+        goals = []
+
+        # Example: synthesize learned concepts into new ideas
+        if self._sufficient_knowledge_for_creativity():
+            goal = Goal(
+                type=GoalType.CREATIVE,
+                description="Synthesize learned concepts into new insights",
+                strategy={
+                    'method': 'concept_synthesis',
+                    'approach': 'analogical_reasoning'
+                },
+                priority=0.4
+            )
+            goals.append(goal)
+
+        return goals
+
+    def prioritize_goals(self, goals: List[Goal]) -> List[Goal]:
+        """
+        Rank goals by priority.
+        Priority based on importance, urgency, and current state.
+        """
+        # Sort by priority
+        return sorted(goals, key=lambda g: g.priority, reverse=True)
+
+    def _should_improve_memory(self) -> bool:
+        """Check if memory system needs improvement"""
+        # Placeholder - would check performance metrics
+        return False
+
+    def _has_interesting_updates(self) -> bool:
+        """Check if have interesting things to share"""
+        # Placeholder - would check recent learning
+        return False
+
+    def _sufficient_knowledge_for_creativity(self) -> bool:
+        """Check if have enough knowledge to be creative"""
+        # Placeholder - would check knowledge base size
+        return False
+
+    def complete_goal(self, goal_id: str):
+        """Mark goal as completed"""
+        for goal in self.active_goals:
+            if goal.id == goal_id:
+                goal.completed = True
+                goal.completed_at = datetime.now(timezone.utc)
+                self.completed_goals.append(goal)
+                self.active_goals.remove(goal)
+                break
+
+    def update_goal_progress(self, goal_id: str, progress: float):
+        """Update progress on a goal"""
+        for goal in self.active_goals:
+            if goal.id == goal_id:
+                goal.progress = min(1.0, progress)
+                if goal.is_complete():
+                    self.complete_goal(goal_id)
+                break
+
+
+class IntentionScheduler:
+    """
+    Decides what to do next - autonomous agency.
+    Executes goals autonomously.
+    """
+
+    def __init__(self, goal_system: GoalFormationSystem):
+        self.goal_system = goal_system
+        self.goal_queue = PriorityQueue()
+        self.current_intention: Optional[Goal] = None
+        self.interrupt_flag = False
+        self.running = False
+
+    async def run_intention_loop(self):
+        """
+        Main autonomous activity loop.
+        Continuously executes highest priority goals.
+        """
+        self.running = True
+
+        while self.running:
+            # Check for interrupts (user interaction)
+            if self.interrupt_flag:
+                await self.handle_interrupt()
+                continue
+
+            # Get highest priority goal
+            if self.current_intention is None:
+                if self.goal_queue.empty():
+                    # Generate new goals
+                    new_goals = self.goal_system.generate_goals()
+                    for goal in new_goals:
+                        self.goal_queue.put(goal)
+
+                    if not self.goal_queue.empty():
+                        self.current_intention = self.goal_queue.get()
+                else:
+                    self.current_intention = self.goal_queue.get()
+
+            # Execute intention
+            if self.current_intention:
+                await self.execute_intention(self.current_intention)
+
+                # Check if complete
+                if self.current_intention.is_complete():
+                    self.goal_system.complete_goal(self.current_intention.id)
+                    self.current_intention = None
+
+            await asyncio.sleep(1)  # Don't busy-wait
+
+    async def execute_intention(self, intention: Goal):
+        """
+        Carry out goal - exploration, learning, conversation, etc.
+        """
+        # Placeholder implementations
+        if intention.type == GoalType.LEARNING:
+            # Would trigger web exploration
+            print(f"Executing learning goal: {intention.description}")
+            intention.progress += 0.1
+
+        elif intention.type == GoalType.SOCIAL:
+            # Would initiate conversation
+            print(f"Executing social goal: {intention.description}")
+            intention.progress += 0.1
+
+        elif intention.type == GoalType.SELF_IMPROVEMENT:
+            # Would enter shop mode
+            print(f"Executing self-improvement goal: {intention.description}")
+            intention.progress += 0.1
+
+        elif intention.type == GoalType.CREATIVE:
+            # Would synthesize concepts
+            print(f"Executing creative goal: {intention.description}")
+            intention.progress += 0.1
+
+    async def handle_interrupt(self):
+        """Handle user interrupt"""
+        # Pause current activity
+        self.interrupt_flag = False
+        # Would transition to conversation mode
+
+    def stop(self):
+        """Stop intention loop"""
+        self.running = False
diff --git a/puma/memory/__init__.py b/puma/memory/__init__.py
new file mode 100644
index 0000000..1d5f56d
--- /dev/null
+++ b/puma/memory/__init__.py
@@ -0,0 +1,10 @@
+"""
+PUMA Memory System
+
+Episodic memory formation, consolidation, and retrieval.
+"""
+
+from .episodic import EpisodicMemorySystem, Episode
+from .consolidation import MemoryConsolidation
+
+__all__ = ['EpisodicMemorySystem', 'Episode', 'MemoryConsolidation']
diff --git a/puma/memory/consolidation.py b/puma/memory/consolidation.py
new file mode 100644
index 0000000..4ba11e6
--- /dev/null
+++ b/puma/memory/consolidation.py
@@ -0,0 +1,278 @@
+"""
+Memory Consolidation
+
+Transforms raw experiences into lasting knowledge during low-activity periods.
+Similar to dreaming - extracts patterns, forms concepts, strengthens important memories.
+"""
+
+from typing import List, Dict, Set, Any
+from collections import Counter, defaultdict
+from dataclasses import dataclass
+import numpy as np
+
+from .episodic import Episode, MemoryType
+
+
+@dataclass
+class Pattern:
+    """Discovered pattern across episodes"""
+    pattern_type: str
+    frequency: int
+    episodes: List[str]
+    description: str
+
+
+@dataclass
+class Concept:
+    """Formed concept from patterns"""
+    name: str
+    definition: str
+    supporting_episodes: List[str]
+    confidence: float
+
+
+class MemoryConsolidation:
+    """
+    Consolidates episodic memories into lasting knowledge.
+    Runs during low-activity periods (sleep mode).
+    """
+
+    def __init__(self, atomspace=None, rft_engine=None):
+        self.atomspace = atomspace
+        self.rft_engine = rft_engine
+        self.patterns_discovered: List[Pattern] = []
+        self.concepts_formed: List[Concept] = []
+
+    async def consolidate(self, episodes: List[Episode]) -> Dict[str, Any]:
+        """
+        Main consolidation process.
+
+        Args:
+            episodes: Batch of episodes to consolidate
+
+        Returns:
+            Dictionary with consolidation results
+        """
+        if not episodes:
+            return {
+                'patterns': [],
+                'concepts': [],
+                'insights': []
+            }
+
+        # Extract patterns across episodes
+        patterns = self.extract_patterns(episodes)
+
+        # Form abstractions (concepts)
+        concepts = self.form_concepts(patterns)
+
+        # Adjust memory weights (strengthen important, fade trivial)
+        self.adjust_memory_weights(episodes)
+
+        # Update self-model based on behavioral patterns
+        self_insights = self.update_self_understanding(patterns)
+
+        # RFT relational frame formation
+        relational_frames = []
+        if self.rft_engine:
+            relational_frames = self.rft_engine.derive_relations(episodes)
+
+        # Store insights in atomspace
+        if self.atomspace:
+            self._store_insights(concepts, relational_frames)
+
+        return {
+            'patterns': patterns,
+            'concepts': concepts,
+            'relational_frames': relational_frames,
+            'insights': self_insights,
+            'episodes_processed': len(episodes)
+        }
+
+    def extract_patterns(self, episodes: List[Episode]) -> List[Pattern]:
+        """
+        Discover patterns in experience.
+        What tends to happen? What co-occurs?
+        """
+        patterns = []
+
+        # Temporal patterns (what follows what)
+        temporal_patterns = self._find_temporal_patterns(episodes)
+        patterns.extend(temporal_patterns)
+
+        # Co-occurrence patterns (what appears together)
+        cooccurrence_patterns = self._find_cooccurrence_patterns(episodes)
+        patterns.extend(cooccurrence_patterns)
+
+        # Outcome patterns (similar outcomes from similar contexts)
+        outcome_patterns = self._find_outcome_patterns(episodes)
+        patterns.extend(outcome_patterns)
+
+        self.patterns_discovered.extend(patterns)
+        return patterns
+
+    def _find_temporal_patterns(self, episodes: List[Episode]) -> List[Pattern]:
+        """Find temporal sequences (A often follows B)"""
+        patterns = []
+        sorted_episodes = sorted(episodes, key=lambda e: e.timestamp)
+
+        # Look for consecutive episode type sequences
+        sequences = []
+        for i in range(len(sorted_episodes) - 1):
+            seq = (sorted_episodes[i].type.value, sorted_episodes[i + 1].type.value)
+            sequences.append(seq)
+
+        # Find frequent sequences
+        seq_counts = Counter(sequences)
+        for seq, count in seq_counts.items():
+            if count >= 2:  # Occurred at least twice
+                patterns.append(Pattern(
+                    pattern_type='temporal',
+                    frequency=count,
+                    episodes=[e.id for e in sorted_episodes],
+                    description=f"{seq[0]} often followed by {seq[1]}"
+                ))
+
+        return patterns
+
+    def _find_cooccurrence_patterns(self, episodes: List[Episode]) -> List[Pattern]:
+        """Find what tends to appear together"""
+        patterns = []
+
+        # Extract context items across episodes
+        context_cooccurrences = defaultdict(int)
+        for episode in episodes:
+            if len(episode.context) > 1:
+                # All pairs in context
+                for i in range(len(episode.context)):
+                    for j in range(i + 1, len(episode.context)):
+                        pair = tuple(sorted([episode.context[i], episode.context[j]]))
+                        context_cooccurrences[pair] += 1
+
+        # Frequent cooccurrences
+        for pair, count in context_cooccurrences.items():
+            if count >= 2:
+                patterns.append(Pattern(
+                    pattern_type='cooccurrence',
+                    frequency=count,
+                    episodes=[e.id for e in episodes if all(p in e.context for p in pair)],
+                    description=f"{pair[0]} and {pair[1]} often occur together"
+                ))
+
+        return patterns
+
+    def _find_outcome_patterns(self, episodes: List[Episode]) -> List[Pattern]:
+        """Find patterns in outcomes"""
+        patterns = []
+
+        # Group episodes by similar outcomes
+        outcome_groups = defaultdict(list)
+        for episode in episodes:
+            if episode.outcome:
+                # Simple grouping by outcome success/failure
+                outcome_key = 'success' if episode.outcome.get('success') else 'failure'
+                outcome_groups[outcome_key].append(episode)
+
+        for outcome_type, group_episodes in outcome_groups.items():
+            if len(group_episodes) >= 2:
+                patterns.append(Pattern(
+                    pattern_type='outcome',
+                    frequency=len(group_episodes),
+                    episodes=[e.id for e in group_episodes],
+                    description=f"Pattern of {outcome_type} outcomes"
+                ))
+
+        return patterns
+
+    def form_concepts(self, patterns: List[Pattern]) -> List[Concept]:
+        """
+        Form abstract concepts from patterns.
+        Concepts are NOT preset - they emerge from experience.
+        """
+        concepts = []
+
+        # Group patterns by type
+        pattern_groups = defaultdict(list)
+        for pattern in patterns:
+            pattern_groups[pattern.pattern_type].append(pattern)
+
+        # Form concepts from pattern groups
+        for pattern_type, group in pattern_groups.items():
+            if len(group) >= 2:
+                concept = Concept(
+                    name=f"{pattern_type}_pattern_concept",
+                    definition=f"Abstract understanding of {pattern_type} patterns",
+                    supporting_episodes=list(set(
+                        ep_id for p in group for ep_id in p.episodes
+                    )),
+                    confidence=min(1.0, len(group) / 10.0)
+                )
+                concepts.append(concept)
+
+        self.concepts_formed.extend(concepts)
+        return concepts
+
+    def adjust_memory_weights(self, episodes: List[Episode]):
+        """
+        Strengthen important memories, fade trivial ones.
+        Importance based on emotional valence, self-relevance, novelty.
+        """
+        for episode in episodes:
+            importance = (
+                abs(episode.emotional_valence) * 0.4 +
+                episode.self_relevance * 0.6
+            )
+
+            # Strengthen important memories in atomspace
+            if self.atomspace and importance > 0.5:
+                # Will implement atomspace strength adjustment
+                pass
+
+            # Fade trivial memories (low importance, not novel)
+            if importance < 0.2:
+                # Mark for potential pruning
+                pass
+
+    def update_self_understanding(self, patterns: List[Pattern]) -> List[str]:
+        """
+        Update self-model based on behavioral patterns.
+        Who am I? What do I tend to do?
+        """
+        insights = []
+
+        # Analyze behavioral tendencies
+        action_patterns = [p for p in patterns if 'action' in p.description.lower()]
+        if action_patterns:
+            insights.append(f"Identified {len(action_patterns)} behavioral tendencies")
+
+        # Analyze emotional patterns
+        emotional_insights = self._analyze_emotional_patterns(patterns)
+        insights.extend(emotional_insights)
+
+        return insights
+
+    def _analyze_emotional_patterns(self, patterns: List[Pattern]) -> List[str]:
+        """Identify emotional tendencies"""
+        insights = []
+
+        # This would analyze emotional valence patterns
+        # For now, placeholder
+        if patterns:
+            insights.append("Emotional pattern analysis complete")
+
+        return insights
+
+    def _store_insights(self, concepts: List[Concept], relational_frames: List):
+        """Store formed concepts and relations in atomspace"""
+        if not self.atomspace:
+            return
+
+        # Store concepts as concept nodes
+        for concept in concepts:
+            # Will implement with actual Atomspace API
+            pass
+
+        # Store relational frames
+        for frame in relational_frames:
+            # Will implement with actual Atomspace API
+            pass
diff --git a/puma/memory/episodic.py b/puma/memory/episodic.py
new file mode 100644
index 0000000..07772f8
--- /dev/null
+++ b/puma/memory/episodic.py
@@ -0,0 +1,202 @@
+"""
+Episodic Memory System
+
+Every experience becomes a memory, shaping identity over time.
+"""
+
+from datetime import datetime, timezone
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Any
+from enum import Enum
+import uuid
+
+
+class MemoryType(Enum):
+    """Types of episodic memories"""
+    CONVERSATION = "conversation"
+    WEB_EXPLORATION = "web_exploration"
+    SELF_MODIFICATION = "self_modification"
+    LEARNING = "learning"
+    GOAL_FORMATION = "goal_formation"
+    STATE_TRANSITION = "state_transition"
+    GENERAL = "general"
+
+
+@dataclass
+class Episode:
+    """
+    Single episodic memory - a moment in time.
+    """
+    id: str = field(default_factory=lambda: str(uuid.uuid4()))
+    timestamp: datetime = field(default_factory=lambda: datetime.now(timezone.utc))
+    type: MemoryType = MemoryType.GENERAL
+    perception: Optional[Dict[str, Any]] = None
+    action: Optional[Dict[str, Any]] = None
+    outcome: Optional[Dict[str, Any]] = None
+    context: List[str] = field(default_factory=list)
+    emotional_valence: float = 0.0
+    self_relevance: float = 0.0
+    consolidated: bool = False
+
+    def to_dict(self) -> Dict:
+        return {
+            'id': self.id,
+            'timestamp': self.timestamp.isoformat(),
+            'type': self.type.value,
+            'perception': self.perception,
+            'action': self.action,
+            'outcome': self.outcome,
+            'context': self.context,
+            'emotional_valence': self.emotional_valence,
+            'self_relevance': self.self_relevance,
+            'consolidated': self.consolidated
+        }
+
+    def generate_summary(self) -> str:
+        """Generate human-readable summary"""
+        if self.type == MemoryType.CONVERSATION:
+            return f"Conversation: {self.perception.get('utterance', 'unknown')[:50]}"
+        elif self.type == MemoryType.WEB_EXPLORATION:
+            return f"Explored: {self.perception.get('url', 'unknown')}"
+        elif self.type == MemoryType.SELF_MODIFICATION:
+            return f"Modified: {self.action.get('module', 'unknown')}"
+        else:
+            return f"{self.type.value} at {self.timestamp.isoformat()}"
+
+
+class EpisodicMemorySystem:
+    """
+    Manages formation and storage of episodic memories.
+    Every experience becomes a memory that shapes identity.
+    """
+
+    def __init__(self, atomspace=None):
+        self.atomspace = atomspace
+        self.episodes: List[Episode] = []
+        self.current_context: List[str] = []
+        self.consolidation_queue: List[Episode] = []
+
+    def form_episode(
+        self,
+        perception: Optional[Dict] = None,
+        action: Optional[Dict] = None,
+        outcome: Optional[Dict] = None,
+        memory_type: MemoryType = MemoryType.GENERAL
+    ) -> Episode:
+        """
+        Create episodic memory from experience.
+
+        Args:
+            perception: What was perceived
+            action: What action was taken
+            outcome: What happened as a result
+            memory_type: Type of memory
+
+        Returns:
+            Episode object
+        """
+        episode = Episode(
+            type=memory_type,
+            perception=perception,
+            action=action,
+            outcome=outcome,
+            context=self.current_context.copy(),
+            emotional_valence=self._assess_emotion(outcome),
+            self_relevance=self._assess_self_relevance(perception, outcome)
+        )
+
+        self.episodes.append(episode)
+        self.consolidation_queue.append(episode)
+
+        # Store in atomspace if available
+        if self.atomspace:
+            self._store_in_atomspace(episode)
+
+        return episode
+
+    def _assess_emotion(self, outcome: Optional[Dict]) -> float:
+        """
+        Assess emotional valence of outcome.
+        Emergent from goal satisfaction, surprise, novelty.
+        NOT preset emotional reactions.
+        """
+        if not outcome:
+            return 0.0
+
+        valence = 0.0
+
+        # Goal satisfaction contributes positive valence
+        if outcome.get('goal_satisfied'):
+            valence += 0.5
+
+        # Surprise can be positive or negative
+        if outcome.get('surprising'):
+            valence += 0.2 if outcome.get('positive_surprise') else -0.2
+
+        # Novelty contributes mild positive valence
+        if outcome.get('novel'):
+            valence += 0.3
+
+        return max(-1.0, min(1.0, valence))
+
+    def _assess_self_relevance(
+        self,
+        perception: Optional[Dict],
+        outcome: Optional[Dict]
+    ) -> float:
+        """
+        Assess how relevant this experience is to self-model.
+        """
+        if not perception and not outcome:
+            return 0.0
+
+        relevance = 0.0
+
+        # Direct self-reference
+        if perception and 'self' in str(perception).lower():
+            relevance += 0.5
+
+        # Goal-relevant
+        if outcome and outcome.get('goal_relevant'):
+            relevance += 0.3
+
+        # Learning-relevant
+        if outcome and outcome.get('learned_something'):
+            relevance += 0.2
+
+        return min(1.0, relevance)
+
+    def _store_in_atomspace(self, episode: Episode):
+        """Store episode in atomspace as node"""
+        # Will integrate with actual Atomspace
+        pass
+
+    def get_recent_episodes(self, limit: int = 10) -> List[Episode]:
+        """Get most recent episodes"""
+        return sorted(self.episodes, key=lambda e: e.timestamp, reverse=True)[:limit]
+
+    def get_unconsolidated_episodes(self) -> List[Episode]:
+        """Get episodes that haven't been consolidated yet"""
+        return [e for e in self.episodes if not e.consolidated]
+
+    def mark_consolidated(self, episode_ids: List[str]):
+        """Mark episodes as consolidated"""
+        for episode in self.episodes:
+            if episode.id in episode_ids:
+                episode.consolidated = True
+
+    def update_context(self, context_items: List[str]):
+        """Update current context for future episodes"""
+        self.current_context = context_items
+
+    def get_episodes_by_type(self, memory_type: MemoryType) -> List[Episode]:
+        """Retrieve episodes by type"""
+        return [e for e in self.episodes if e.type == memory_type]
+
+    def count_total_episodes(self) -> int:
+        """Count all episodes"""
+        return len(self.episodes)
+
+    def get_timeline(self) -> List[Episode]:
+        """Get full autobiographical timeline"""
+        return sorted(self.episodes, key=lambda e: e.timestamp)
diff --git a/puma/rft/__init__.py b/puma/rft/__init__.py
index 43fbe0c..bd0d538 100644
--- a/puma/rft/__init__.py
+++ b/puma/rft/__init__.py
@@ -13,6 +13,7 @@
 )
 from .orchestrator import solve_with_rft, insert_after_stuck_rule
 from .explain import explain_last_run, record_trace
+from .reasoning import RFTEngine, RelationType, RelationalFrame
 
 __all__ = [
     "RFT_ENABLED",
@@ -34,4 +35,7 @@
     "insert_after_stuck_rule",
     "explain_last_run",
     "record_trace",
+    "RFTEngine",
+    "RelationType",
+    "RelationalFrame",
 ]
diff --git a/puma/rft/reasoning.py b/puma/rft/reasoning.py
new file mode 100644
index 0000000..db12475
--- /dev/null
+++ b/puma/rft/reasoning.py
@@ -0,0 +1,354 @@
+"""
+RFT Reasoning Engine
+
+Implements Relational Frame Theory for analogical reasoning and relation derivation.
+Learns arbitrary relations and enables derivation of new relations without explicit training.
+"""
+
+from typing import List, Dict, Optional, Any, Tuple
+from dataclasses import dataclass
+from enum import Enum
+import numpy as np
+from collections import defaultdict
+
+
+class RelationType(Enum):
+    """Types of relational frames"""
+    COORDINATION = "coordination"  # Similarity (X is like Y)
+    OPPOSITION = "opposition"  # Difference (X is opposite of Y)
+    HIERARCHY = "hierarchy"  # Categorization (X is a type of Y)
+    TEMPORAL = "temporal"  # Before/after (X happens before Y)
+    CAUSAL = "causal"  # If-then (X causes Y)
+    COMPARATIVE = "comparative"  # More/less (X is more than Y)
+    SPATIAL = "spatial"  # Location (X is near Y)
+
+
+@dataclass
+class RelationalFrame:
+    """
+    A learned relational frame between stimuli.
+    Enables derivation: if A relates to B, and B relates to C,
+    then A relates to C (transitivity).
+    """
+    relation_type: RelationType
+    source: str
+    target: str
+    strength: float
+    context: Optional[List[str]] = None
+    derived: bool = False  # Was this derived vs. directly learned?
+
+    def to_dict(self) -> Dict:
+        return {
+            'relation_type': self.relation_type.value,
+            'source': self.source,
+            'target': self.target,
+            'strength': self.strength,
+            'context': self.context,
+            'derived': self.derived
+        }
+
+
+class RFTEngine:
+    """
+    Relational Frame Theory reasoning engine.
+    Learns relational frames and derives new relations through transformation.
+    """
+
+    def __init__(self, atomspace=None):
+        self.atomspace = atomspace
+        self.frames: List[RelationalFrame] = []
+        self.frame_index: Dict[str, List[RelationalFrame]] = defaultdict(list)
+
+    def derive_relations(self, episodes: List) -> List[RelationalFrame]:
+        """
+        Discover relational patterns in episodes.
+
+        Args:
+            episodes: List of Episode objects
+
+        Returns:
+            List of discovered relational frames
+        """
+        relations = []
+
+        # Find coordination frames (similarity)
+        similarities = self.find_coordination_frames(episodes)
+        relations.extend(similarities)
+
+        # Find opposition frames
+        oppositions = self.find_opposition_frames(episodes)
+        relations.extend(oppositions)
+
+        # Find hierarchical relations
+        hierarchies = self.find_hierarchy_frames(episodes)
+        relations.extend(hierarchies)
+
+        # Find temporal relations
+        temporal = self.find_temporal_frames(episodes)
+        relations.extend(temporal)
+
+        # Find causal relations
+        causal = self.find_causal_frames(episodes)
+        relations.extend(causal)
+
+        # Store frames
+        for frame in relations:
+            self.add_frame(frame)
+
+        return relations
+
+    def find_coordination_frames(self, episodes: List) -> List[RelationalFrame]:
+        """
+        Find similarity relations (X is like Y).
+        Basis for analogical reasoning.
+        """
+        frames = []
+
+        # Group episodes by similar outcomes
+        outcome_groups = defaultdict(list)
+        for episode in episodes:
+            if episode.outcome:
+                # Simple grouping by success/failure
+                outcome_key = 'success' if episode.outcome.get('success') else 'failure'
+                outcome_groups[outcome_key].append(episode)
+
+        # Episodes in same group are similar
+        for group_episodes in outcome_groups.values():
+            if len(group_episodes) >= 2:
+                # Create coordination frames between similar episodes
+                for i in range(len(group_episodes) - 1):
+                    frame = RelationalFrame(
+                        relation_type=RelationType.COORDINATION,
+                        source=group_episodes[i].id,
+                        target=group_episodes[i + 1].id,
+                        strength=0.7,
+                        context=['similar_outcome']
+                    )
+                    frames.append(frame)
+
+        return frames
+
+    def find_opposition_frames(self, episodes: List) -> List[RelationalFrame]:
+        """
+        Find opposition relations (X is opposite of Y).
+        """
+        frames = []
+
+        # Find episodes with opposite emotional valences
+        positive_episodes = [e for e in episodes if e.emotional_valence > 0.5]
+        negative_episodes = [e for e in episodes if e.emotional_valence < -0.5]
+
+        # Create opposition frames
+        for pos_ep in positive_episodes:
+            for neg_ep in negative_episodes:
+                frame = RelationalFrame(
+                    relation_type=RelationType.OPPOSITION,
+                    source=pos_ep.id,
+                    target=neg_ep.id,
+                    strength=0.6,
+                    context=['opposite_valence']
+                )
+                frames.append(frame)
+                break  # Only create one example per positive episode
+
+        return frames
+
+    def find_hierarchy_frames(self, episodes: List) -> List[RelationalFrame]:
+        """
+        Find hierarchical relations (X is a type of Y).
+        """
+        frames = []
+
+        # Group episodes by type
+        type_groups = defaultdict(list)
+        for episode in episodes:
+            type_groups[episode.type.value].append(episode)
+
+        # All episodes of a type are instances of that type
+        for episode_type, group_episodes in type_groups.items():
+            for episode in group_episodes:
+                frame = RelationalFrame(
+                    relation_type=RelationType.HIERARCHY,
+                    source=episode.id,
+                    target=f"category:{episode_type}",
+                    strength=1.0,
+                    context=['type_hierarchy']
+                )
+                frames.append(frame)
+
+        return frames
+
+    def find_temporal_frames(self, episodes: List) -> List[RelationalFrame]:
+        """
+        Find temporal relations (X before Y).
+        """
+        frames = []
+
+        # Sort by timestamp
+        sorted_episodes = sorted(episodes, key=lambda e: e.timestamp)
+
+        # Create temporal frames for consecutive episodes
+        for i in range(len(sorted_episodes) - 1):
+            frame = RelationalFrame(
+                relation_type=RelationType.TEMPORAL,
+                source=sorted_episodes[i].id,
+                target=sorted_episodes[i + 1].id,
+                strength=1.0,  # Temporal order is definite
+                context=['chronological']
+            )
+            frames.append(frame)
+
+        return frames
+
+    def find_causal_frames(self, episodes: List) -> List[RelationalFrame]:
+        """
+        Find causal relations (X causes Y).
+        If action in episode N, outcome in episode N+1 -> potential causation.
+        """
+        frames = []
+
+        sorted_episodes = sorted(episodes, key=lambda e: e.timestamp)
+
+        for i in range(len(sorted_episodes) - 1):
+            current = sorted_episodes[i]
+            next_ep = sorted_episodes[i + 1]
+
+            # If current has action and next has outcome, potential causation
+            if current.action and next_ep.outcome:
+                frame = RelationalFrame(
+                    relation_type=RelationType.CAUSAL,
+                    source=current.id,
+                    target=next_ep.id,
+                    strength=0.5,  # Uncertain - correlation not causation
+                    context=['potential_causation']
+                )
+                frames.append(frame)
+
+        return frames
+
+    def add_frame(self, frame: RelationalFrame):
+        """Add frame to knowledge base"""
+        self.frames.append(frame)
+        self.frame_index[frame.source].append(frame)
+        self.frame_index[frame.target].append(frame)
+
+    def derive_by_transitivity(
+        self,
+        relation_type: RelationType,
+        start: str,
+        end: str
+    ) -> Optional[RelationalFrame]:
+        """
+        Derive new relation through transitivity.
+        If A->B and B->C, then A->C (for transitive relations).
+        """
+        # Find path from start to end
+        path = self._find_relation_path(start, end, relation_type)
+
+        if path and len(path) >= 2:
+            # Can derive relation
+            # Strength decays with path length
+            strength = 1.0 / len(path)
+
+            derived_frame = RelationalFrame(
+                relation_type=relation_type,
+                source=start,
+                target=end,
+                strength=strength,
+                derived=True,
+                context=['derived_by_transitivity']
+            )
+
+            return derived_frame
+
+        return None
+
+    def _find_relation_path(
+        self,
+        start: str,
+        end: str,
+        relation_type: RelationType,
+        max_depth: int = 3
+    ) -> Optional[List[str]]:
+        """
+        Find path of relations from start to end.
+        Simple BFS.
+        """
+        if start == end:
+            return [start]
+
+        visited = set()
+        queue = [(start, [start])]
+
+        while queue and len(queue[0][1]) <= max_depth:
+            current, path = queue.pop(0)
+
+            if current in visited:
+                continue
+
+            visited.add(current)
+
+            # Get outgoing frames
+            for frame in self.frame_index.get(current, []):
+                if frame.relation_type == relation_type and frame.source == current:
+                    new_path = path + [frame.target]
+
+                    if frame.target == end:
+                        return new_path
+
+                    queue.append((frame.target, new_path))
+
+        return None
+
+    def reason_by_analogy(self, novel_situation: Dict[str, Any]) -> Dict[str, Any]:
+        """
+        Use past experience to understand new situations.
+        Find similar past episodes via relational frames.
+        """
+        # Find similar episodes using coordination frames
+        similar_frames = [
+            f for f in self.frames
+            if f.relation_type == RelationType.COORDINATION
+        ]
+
+        if not similar_frames:
+            return {'prediction': None, 'confidence': 0.0}
+
+        # For now, simple heuristic - return most recent similar frame
+        most_recent = max(similar_frames, key=lambda f: f.strength)
+
+        return {
+            'prediction': f"Similar to {most_recent.source}",
+            'confidence': most_recent.strength,
+            'analogy_source': most_recent.source
+        }
+
+    def get_related_items(
+        self,
+        item: str,
+        relation_type: Optional[RelationType] = None
+    ) -> List[RelationalFrame]:
+        """
+        Get all items related to given item.
+        """
+        frames = self.frame_index.get(item, [])
+
+        if relation_type:
+            frames = [f for f in frames if f.relation_type == relation_type]
+
+        return frames
+
+    def get_frame_statistics(self) -> Dict[str, int]:
+        """Get statistics about learned frames"""
+        stats = defaultdict(int)
+
+        for frame in self.frames:
+            stats[frame.relation_type.value] += 1
+            if frame.derived:
+                stats['derived'] += 1
+            else:
+                stats['learned'] += 1
+
+        stats['total'] = len(self.frames)
+
+        return dict(stats)
diff --git a/puma/shop/__init__.py b/puma/shop/__init__.py
new file mode 100644
index 0000000..1214959
--- /dev/null
+++ b/puma/shop/__init__.py
@@ -0,0 +1,18 @@
+"""
+The Shop - Self-Modification System
+
+AGI can inspect and modify its own cognitive code.
+"""
+
+from .introspection import CodeIntrospection, CognitiveModule
+from .modification import ModificationSystem, ModificationPlan
+from .sandbox import ModificationSandbox, TestReport
+
+__all__ = [
+    'CodeIntrospection',
+    'CognitiveModule',
+    'ModificationSystem',
+    'ModificationPlan',
+    'ModificationSandbox',
+    'TestReport'
+]
diff --git a/puma/shop/introspection.py b/puma/shop/introspection.py
new file mode 100644
index 0000000..5673d47
--- /dev/null
+++ b/puma/shop/introspection.py
@@ -0,0 +1,181 @@
+"""
+Code Introspection
+
+AGI builds internal model of its own code and assesses performance.
+"""
+
+import ast
+import inspect
+from pathlib import Path
+from typing import Dict, List, Optional, Any
+from dataclasses import dataclass
+from collections import defaultdict
+
+
+@dataclass
+class CognitiveModule:
+    """Internal representation of a cognitive code module"""
+    name: str
+    path: Path
+    functions: List[str]
+    classes: List[str]
+    dependencies: List[str]
+    purpose: str
+    lines_of_code: int
+
+
+class CodeIntrospection:
+    """
+    Builds internal model of cognitive architecture code.
+    """
+
+    def __init__(self, codebase_path: Path):
+        self.codebase_path = codebase_path
+        self.cognitive_modules: Dict[str, CognitiveModule] = {}
+        self.modification_history: List[Dict] = []
+
+    def map_cognitive_architecture(self) -> Dict[str, CognitiveModule]:
+        """
+        Build internal model of own code.
+        """
+        modules = {
+            'perception': self.analyze_module(self.codebase_path / 'gemini-interface'),
+            'memory': self.analyze_module(self.codebase_path / 'puma' / 'memory'),
+            'reasoning': self.analyze_module(self.codebase_path / 'puma' / 'rft'),
+            'goals': self.analyze_module(self.codebase_path / 'puma' / 'goals'),
+            'curiosity': self.analyze_module(self.codebase_path / 'puma' / 'curiosity'),
+            'shop': self.analyze_module(self.codebase_path / 'puma' / 'shop'),
+        }
+
+        self.cognitive_modules = modules
+        return modules
+
+    def analyze_module(self, module_path: Path) -> CognitiveModule:
+        """
+        Parse code module into understanding.
+        """
+        if not module_path.exists():
+            return CognitiveModule(
+                name=module_path.name,
+                path=module_path,
+                functions=[],
+                classes=[],
+                dependencies=[],
+                purpose="Not yet implemented",
+                lines_of_code=0
+            )
+
+        functions = []
+        classes = []
+        dependencies = []
+        total_lines = 0
+
+        # Analyze all Python files in module
+        for py_file in module_path.rglob('*.py'):
+            if py_file.name.startswith('__'):
+                continue
+
+            try:
+                with open(py_file, 'r') as f:
+                    source = f.read()
+                    total_lines += len(source.split('\n'))
+
+                # Parse AST
+                tree = ast.parse(source)
+
+                # Extract functions
+                for node in ast.walk(tree):
+                    if isinstance(node, ast.FunctionDef):
+                        functions.append(node.name)
+                    elif isinstance(node, ast.ClassDef):
+                        classes.append(node.name)
+                    elif isinstance(node, ast.Import):
+                        for alias in node.names:
+                            dependencies.append(alias.name)
+                    elif isinstance(node, ast.ImportFrom):
+                        if node.module:
+                            dependencies.append(node.module)
+
+            except Exception as e:
+                # Skip files that can't be parsed
+                pass
+
+        # Infer purpose from name and structure
+        purpose = self._infer_module_purpose(module_path.name, functions, classes)
+
+        return CognitiveModule(
+            name=module_path.name,
+            path=module_path,
+            functions=list(set(functions)),
+            classes=list(set(classes)),
+            dependencies=list(set(dependencies)),
+            purpose=purpose,
+            lines_of_code=total_lines
+        )
+
+    def _infer_module_purpose(
+        self,
+        module_name: str,
+        functions: List[str],
+        classes: List[str]
+    ) -> str:
+        """
+        Infer module purpose from name and structure.
+        """
+        purposes = {
+            'memory': 'Episodic memory and consolidation',
+            'rft': 'Relational reasoning and analogical thinking',
+            'goals': 'Goal formation and intention scheduling',
+            'curiosity': 'Intrinsic motivation and question generation',
+            'shop': 'Self-modification and code introspection',
+            'consciousness': 'State management and coordination',
+        }
+
+        return purposes.get(module_name, f"Module: {module_name}")
+
+    def assess_cognitive_performance(self) -> Dict[str, float]:
+        """
+        Identify areas for self-improvement.
+        Measures cognitive performance metrics.
+        """
+        metrics = {
+            'memory_efficiency': self.measure_memory_performance(),
+            'reasoning_speed': self.measure_reasoning_speed(),
+            'learning_rate': self.measure_learning_effectiveness(),
+            'goal_completion': self.measure_goal_success_rate(),
+            'curiosity_satisfaction': self.measure_question_resolution()
+        }
+
+        return metrics
+
+    def measure_memory_performance(self) -> float:
+        """Measure memory system performance"""
+        # Placeholder - would measure consolidation speed, retrieval accuracy
+        return 0.7
+
+    def measure_reasoning_speed(self) -> float:
+        """Measure reasoning engine speed"""
+        # Placeholder - would measure inference time
+        return 0.6
+
+    def measure_learning_effectiveness(self) -> float:
+        """Measure how effectively new knowledge is acquired"""
+        # Placeholder - would measure concept formation rate
+        return 0.8
+
+    def measure_goal_success_rate(self) -> float:
+        """Measure goal completion rate"""
+        # Placeholder - would check goal completion stats
+        return 0.75
+
+    def measure_question_resolution(self) -> float:
+        """Measure how many curiosity questions get answered"""
+        # Placeholder - would check question answer rate
+        return 0.65
+
+    def identify_bottlenecks(self, metrics: Dict[str, float], threshold: float = 0.7) -> Dict[str, float]:
+        """
+        Identify performance bottlenecks.
+        """
+        bottlenecks = {k: v for k, v in metrics.items() if v < threshold}
+        return bottlenecks
diff --git a/puma/shop/modification.py b/puma/shop/modification.py
new file mode 100644
index 0000000..48a6d77
--- /dev/null
+++ b/puma/shop/modification.py
@@ -0,0 +1,200 @@
+"""
+Modification Planning and Execution
+
+Plans and implements code changes to improve cognitive abilities.
+"""
+
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Any
+from pathlib import Path
+from datetime import datetime, timezone
+
+
+@dataclass
+class ModificationPlan:
+    """Plan for self-modification"""
+    target_module: str
+    improvement_goal: str
+    current_code: str
+    proposed_code: str
+    hypothesis: str
+    test_strategy: Dict[str, Any]
+    risk_level: str  # 'low', 'medium', 'high'
+    requires_approval: bool = True
+
+
+class ModificationSystem:
+    """
+    Plans and executes self-modifications.
+    """
+
+    def __init__(self, introspection, atomspace=None):
+        self.introspection = introspection
+        self.atomspace = atomspace
+        self.modification_queue: List[ModificationPlan] = []
+        self.pending_approval: List[ModificationPlan] = []
+
+    def plan_self_modification(self, improvement_goal: Dict[str, Any]) -> ModificationPlan:
+        """
+        Design code changes to improve cognitive ability.
+
+        Args:
+            improvement_goal: Dict with target_module, performance_target
+
+        Returns:
+            ModificationPlan
+        """
+        target_module = improvement_goal['target_module']
+        desired_improvement = improvement_goal['performance_target']
+
+        # Analyze current implementation
+        current_module = self.introspection.cognitive_modules.get(target_module)
+
+        if not current_module:
+            raise ValueError(f"Module {target_module} not found")
+
+        # Read current code
+        current_code = self._read_module_code(current_module)
+
+        # Generate modification hypotheses
+        hypotheses = self.generate_modification_ideas(
+            current_module,
+            desired_improvement
+        )
+
+        # Rank by expected impact
+        best_hypothesis = hypotheses[0] if hypotheses else "Optimize performance"
+
+        # Create modification plan
+        plan = ModificationPlan(
+            target_module=target_module,
+            improvement_goal=desired_improvement,
+            current_code=current_code,
+            proposed_code="# Generated through Gemini collaboration",
+            hypothesis=best_hypothesis,
+            test_strategy=self.design_test(target_module, best_hypothesis),
+            risk_level=self.assess_risk_level(target_module),
+            requires_approval=self.assess_risk_level(target_module) in ['medium', 'high']
+        )
+
+        return plan
+
+    def _read_module_code(self, module) -> str:
+        """Read module source code"""
+        code_parts = []
+
+        if module.path.exists():
+            for py_file in module.path.rglob('*.py'):
+                try:
+                    with open(py_file, 'r') as f:
+                        code_parts.append(f"# File: {py_file.name}\n{f.read()}\n")
+                except:
+                    pass
+
+        return "\n".join(code_parts)
+
+    def generate_modification_ideas(
+        self,
+        module,
+        desired_improvement: str
+    ) -> List[str]:
+        """
+        Generate hypotheses for how to improve module.
+        """
+        ideas = []
+
+        if 'speed' in desired_improvement.lower():
+            ideas.append("Optimize algorithm for faster execution")
+            ideas.append("Add caching to reduce redundant computation")
+            ideas.append("Parallelize independent operations")
+
+        if 'memory' in desired_improvement.lower():
+            ideas.append("Implement more efficient data structures")
+            ideas.append("Add garbage collection optimization")
+
+        if 'accuracy' in desired_improvement.lower():
+            ideas.append("Improve pattern recognition algorithms")
+            ideas.append("Add validation and error checking")
+
+        # Default
+        if not ideas:
+            ideas.append("General performance optimization")
+
+        return ideas
+
+    def design_test(self, module_name: str, hypothesis: str) -> Dict[str, Any]:
+        """
+        Design test strategy for modification.
+        """
+        return {
+            'test_type': 'A/B comparison',
+            'metrics': ['execution_time', 'accuracy', 'memory_usage'],
+            'test_scenarios': [
+                'typical_workload',
+                'edge_cases',
+                'stress_test'
+            ],
+            'success_criteria': {
+                'performance_improvement': 0.2,  # 20% improvement
+                'no_regressions': True
+            }
+        }
+
+    def assess_risk_level(self, module_name: str) -> str:
+        """
+        Assess risk of modifying module.
+        Core modules are high risk, peripheral modules are low risk.
+        """
+        high_risk_modules = ['memory', 'consciousness', 'atomspace']
+        medium_risk_modules = ['rft', 'goals']
+
+        if module_name in high_risk_modules:
+            return 'high'
+        elif module_name in medium_risk_modules:
+            return 'medium'
+        else:
+            return 'low'
+
+    async def request_modification_approval(
+        self,
+        modification_plan: ModificationPlan
+    ) -> bool:
+        """
+        Ask user for approval before self-modifying.
+        Returns True if approved, False if rejected.
+        """
+        # Would integrate with GUI to show approval dialog
+        # For now, placeholder
+        self.pending_approval.append(modification_plan)
+
+        # In real implementation, would await user response
+        # For now, automatically approve low-risk modifications
+        if modification_plan.risk_level == 'low':
+            return True
+
+        return False  # Wait for explicit approval
+
+    def approve_modification(self, plan_id: str):
+        """Approve pending modification"""
+        # Find and approve modification
+        pass
+
+    def log_modification(self, plan: ModificationPlan, outcome: Dict):
+        """
+        Record modification to autobiographical memory.
+        """
+        modification_record = {
+            'timestamp': datetime.now(timezone.utc).isoformat(),
+            'module': plan.target_module,
+            'goal': plan.improvement_goal,
+            'hypothesis': plan.hypothesis,
+            'outcome': outcome,
+            'risk_level': plan.risk_level
+        }
+
+        self.introspection.modification_history.append(modification_record)
+
+        # Store in atomspace episodic memory
+        if self.atomspace:
+            # Would create episodic memory node
+            pass
diff --git a/puma/shop/sandbox.py b/puma/shop/sandbox.py
new file mode 100644
index 0000000..6b88be8
--- /dev/null
+++ b/puma/shop/sandbox.py
@@ -0,0 +1,145 @@
+"""
+Sandboxed Testing
+
+Test self-modifications safely before applying to production.
+"""
+
+from dataclasses import dataclass
+from typing import Dict, Any, List
+from pathlib import Path
+import tempfile
+import shutil
+
+
+@dataclass
+class TestReport:
+    """Results from sandbox testing"""
+    success: bool
+    performance_delta: Dict[str, float]
+    side_effects: List[str]
+    recommendation: str
+    detailed_results: Dict[str, Any]
+
+
+class ModificationSandbox:
+    """
+    Test self-modifications in isolation before deployment.
+    """
+
+    def __init__(self):
+        self.sandbox_dir: Optional[Path] = None
+        self.test_results: List[TestReport] = []
+
+    async def test_modification(self, modification_plan) -> TestReport:
+        """
+        Run modified code in isolated environment.
+
+        Args:
+            modification_plan: ModificationPlan to test
+
+        Returns:
+            TestReport with results
+        """
+        # Create sandbox
+        self.sandbox_dir = Path(tempfile.mkdtemp(prefix='puma_sandbox_'))
+
+        try:
+            # Copy architecture to sandbox
+            sandbox_copy = self._copy_architecture_to_sandbox()
+
+            # Apply modification
+            self._apply_modification(modification_plan)
+
+            # Run test suite
+            results = await self._run_tests(modification_plan.test_strategy)
+
+            # Evaluate results
+            success = self._evaluate_test_results(results)
+
+            report = TestReport(
+                success=success,
+                performance_delta=results.get('performance_delta', {}),
+                side_effects=results.get('side_effects', []),
+                recommendation='approve' if success else 'reject',
+                detailed_results=results
+            )
+
+            self.test_results.append(report)
+            return report
+
+        finally:
+            # Cleanup sandbox
+            if self.sandbox_dir and self.sandbox_dir.exists():
+                shutil.rmtree(self.sandbox_dir)
+
+    def _copy_architecture_to_sandbox(self) -> Path:
+        """Copy cognitive architecture to sandbox"""
+        # Would copy relevant modules
+        return self.sandbox_dir
+
+    def _apply_modification(self, plan):
+        """Apply modification in sandbox"""
+        # Would write modified code to sandbox
+        pass
+
+    async def _run_tests(self, test_strategy: Dict) -> Dict[str, Any]:
+        """Execute test suite"""
+        results = {
+            'performance_delta': {
+                'execution_time': -0.15,  # 15% faster (example)
+                'memory_usage': 0.05,     # 5% more memory
+                'accuracy': 0.0           # No change
+            },
+            'side_effects': [],
+            'all_tests_passed': True
+        }
+
+        # Would run actual tests
+        return results
+
+    def _evaluate_test_results(self, results: Dict) -> bool:
+        """Evaluate if modification is successful"""
+        # Check if all tests passed
+        if not results.get('all_tests_passed', False):
+            return False
+
+        # Check if performance improved
+        perf_delta = results.get('performance_delta', {})
+        if perf_delta.get('execution_time', 0) > 0:  # Slower
+            return False
+
+        # Check for side effects
+        if results.get('side_effects'):
+            return False
+
+        return True
+
+    async def run_cognitive_ab_test(self, original, modified) -> Dict:
+        """
+        Compare cognitive performance: original vs modified.
+        """
+        test_scenarios = self._generate_test_scenarios()
+
+        results_original = []
+        results_modified = []
+
+        for scenario in test_scenarios:
+            # Would test both versions
+            # Placeholder for now
+            pass
+
+        comparison = {
+            'winner': 'modified',
+            'improvement': 0.2,
+            'confidence': 0.85
+        }
+
+        return comparison
+
+    def _generate_test_scenarios(self) -> List[Dict]:
+        """Generate test scenarios for comparison"""
+        return [
+            {'type': 'typical_workload'},
+            {'type': 'edge_case'},
+            {'type': 'stress_test'}
+        ]
diff --git a/requirements.txt b/requirements.txt
index cf8caa5..f240408 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,3 +1,29 @@
+# Core dependencies
 numpy==1.26.4
 hypothesis==6.100.2
 scipy==1.12.0
+
+# Cognitive architecture
+# hyperon will be built from source
+
+# API integrations
+google-generativeai  # Gemini API
+
+# Web agent (optional - for autonomous browsing)
+# playwright  # Autonomous browsing
+# beautifulsoup4  # HTML parsing
+aiohttp  # Async HTTP
+
+# Backend server
+fastapi  # WebSocket server
+uvicorn  # ASGI server
+python-multipart  # File uploads
+websockets  # WebSocket support
+
+# Persistence (optional backends)
+# rocksdb-python  # RocksDB backend
+# psycopg2-binary  # PostgreSQL backend
+
+# Testing
+pytest
+pytest-asyncio
diff --git a/setup_persistence.py b/setup_persistence.py
new file mode 100644
index 0000000..a659ac1
--- /dev/null
+++ b/setup_persistence.py
@@ -0,0 +1,53 @@
+"""
+Setup Persistence Database
+
+Initialize atomspace persistence directory.
+"""
+
+from pathlib import Path
+import json
+
+
+def setup_persistence(base_path: Path = Path("./atomspace-db")):
+    """
+    Set up persistence directory structure.
+
+    Args:
+        base_path: Base path for atomspace database
+    """
+    base_path = Path(base_path)
+
+    # Create directories
+    (base_path / "default").mkdir(parents=True, exist_ok=True)
+    (base_path / "default" / "snapshots").mkdir(exist_ok=True)
+
+    # Create initial empty atoms file
+    atoms_file = base_path / "default" / "atoms.json"
+    if not atoms_file.exists():
+        with open(atoms_file, 'w') as f:
+            json.dump({}, f)
+
+    # Create initial empty links file
+    links_file = base_path / "default" / "links.json"
+    if not links_file.exists():
+        with open(links_file, 'w') as f:
+            json.dump([], f)
+
+    # Create config
+    config_file = base_path / "config.json"
+    config = {
+        "version": "1.0",
+        "default_db": "default",
+        "auto_save_interval": 60,
+        "snapshot_interval": 3600
+    }
+
+    with open(config_file, 'w') as f:
+        json.dump(config, f, indent=2)
+
+    print(f"✅ Persistence setup complete at: {base_path}")
+    print(f"   Default database: {base_path / 'default'}")
+
+
+if __name__ == "__main__":
+    setup_persistence()
diff --git a/web-agent/__init__.py b/web-agent/__init__.py
new file mode 100644
index 0000000..5679de6
--- /dev/null
+++ b/web-agent/__init__.py
@@ -0,0 +1,9 @@
+"""
+Web Agent
+
+Autonomous web browsing and learning.
+"""
+
+from .agent import AutonomousWebAgent
+
+__all__ = ['AutonomousWebAgent']
diff --git a/web-agent/agent.py b/web-agent/agent.py
new file mode 100644
index 0000000..620b918
--- /dev/null
+++ b/web-agent/agent.py
@@ -0,0 +1,156 @@
+"""
+Autonomous Web Agent
+
+AGI decides what to learn, not hardcoded curriculum.
+Browses web autonomously to answer curiosity questions.
+"""
+
+from typing import Dict, List, Optional, Any
+from datetime import datetime, timezone
+from dataclasses import dataclass
+
+
+@dataclass
+class WebExploration:
+    """Record of web exploration"""
+    url: str
+    timestamp: datetime
+    learned_concepts: List[str]
+    questions_answered: List[str]
+    new_questions: List[str]
+
+
+class AutonomousWebAgent:
+    """
+    Autonomous web browsing for learning.
+    """
+
+    def __init__(self, consciousness=None):
+        self.consciousness = consciousness
+        self.browser = None  # Would be Playwright browser
+        self.exploration_history: List[WebExploration] = []
+
+    async def initialize_browser(self):
+        """Initialize headless browser"""
+        print("🌐 Initializing browser...")
+
+        # Would initialize Playwright
+        # from playwright.async_api import async_playwright
+        # playwright = await async_playwright().start()
+        # self.browser = await playwright.chromium.launch()
+
+        self.browser = "placeholder_browser"
+
+    async def explore_from_curiosity(self, curiosity_question: str):
+        """
+        AGI decides what to learn, not hardcoded curriculum.
+
+        Args:
+            curiosity_question: Question from curiosity drive
+        """
+        print(f"🔍 Exploring to answer: {curiosity_question}")
+
+        # Generate search query from internal question
+        query = await self.formulate_search_query(curiosity_question)
+
+        # Search and select pages
+        results = await self.search(query)
+        interesting_pages = await self.select_interesting_results(results)
+
+        # Browse and learn
+        for page_url in interesting_pages:
+            learning = await self.browse_and_extract(page_url)
+
+            if self.consciousness:
+                await self.consciousness.integrate_learning(learning)
+
+    async def formulate_search_query(self, question: str) -> str:
+        """Turn curiosity question into search query"""
+        # Simple implementation - in practice, would use NLP
+        # Remove question words
+        query = question.lower()
+        for word in ['what', 'is', 'how', 'why', 'when', 'where']:
+            query = query.replace(word, '')
+
+        return query.strip()
+
+    async def search(self, query: str) -> List[Dict]:
+        """
+        Perform web search.
+        Returns search results.
+        """
+        print(f"🔎 Searching for: {query}")
+
+        # Placeholder - would use actual search API
+        # In real implementation, use DuckDuckGo or similar
+        results = [
+            {'title': f'Result for {query}', 'url': f'https://example.com/{query}'}
+        ]
+
+        return results
+
+    async def select_interesting_results(self, results: List[Dict]) -> List[str]:
+        """
+        AGI chooses which results to explore.
+        """
+        # Simple implementation - take top 3
+        # In practice, would use relevance scoring
+        return [r['url'] for r in results[:3]]
+
+    async def browse_and_extract(self, url: str) -> WebExploration:
+        """
+        Extract knowledge from page, not just text.
+        """
+        print(f"📖 Browsing: {url}")
+
+        # Placeholder - would use actual browser
+        # page = await self.browser.new_page()
+        # await page.goto(url)
+        # content = await page.content()
+
+        # Parse meaningfully
+        learned_concepts = ['placeholder_concept']
+        questions_answered = []
+        new_questions = ['What else relates to this?']
+
+        exploration = WebExploration(
+            url=url,
+            timestamp=datetime.now(timezone.utc),
+            learned_concepts=learned_concepts,
+            questions_answered=questions_answered,
+            new_questions=new_questions
+        )
+
+        self.exploration_history.append(exploration)
+        return exploration
+
+    async def integrate_web_learning(self, exploration: WebExploration, memory_system):
+        """
+        Web experiences become part of autobiographical memory.
+        """
+        if not memory_system:
+            return
+
+        # Create episodic memory of exploration
+        episode = memory_system.form_episode(
+            perception={
+                'type': 'web_exploration',
+                'url': exploration.url
+            },
+            action={
+                'type': 'browse_and_learn'
+            },
+            outcome={
+                'learned_concepts': exploration.learned_concepts,
+                'questions_answered': exploration.questions_answered,
+                'new_questions': exploration.new_questions
+            },
+            memory_type='web_exploration'
+        )
+
+        return episode
+
+    async def close(self):
+        """Close browser"""
+        if self.browser and self.browser != "placeholder_browser":
+            await self.browser.close()