Skip to content

Latest commit

 

History

History
93 lines (68 loc) · 5.55 KB

CHANGELOG.md

File metadata and controls

93 lines (68 loc) · 5.55 KB

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.6.1] - 2026-03-03

Added

  • Sentence-level memory splitting — Multi-sentence turns are split into individual sentences, each evaluated and saved independently with its own category and keywords
    • split_sentences() splits on sentence-ending punctuation and Korean 종결어미+comma
    • Both real-time (auto_search) and batch (sleep cycle) extraction use sentence splitting
    • Example: "저는 Python 개발자입니다. 김치찌개를 좋아합니다." → [technical] + [preference]

Changed

  • HeuristicMemoryExtractor.evaluate() minimum content length lowered from 20 to 5 chars (Korean text carries high info density per character)

[0.6.0] - 2026-03-03

Added

  • Automatic conversation logging — All conversation turns are recorded to a SQLite append-only log (conversation_log.db) with WAL mode for concurrent safety
  • Dual-saving in auto_search — Every turn is logged to SQLite for batch processing, and high-value turns (matching personal/preference/tech/emotion patterns) are instantly extracted to ChromaDB
  • Memory extraction pipeline — New extraction.py module with three extractors:
    • HeuristicMemoryExtractor — Pattern-matching based (reuses extract_keywords() / classify_category())
    • RLMemoryExtractor — MLP bandit for EXTRACT/SKIP binary decisions with imitation learning
    • ProgressiveExtraction — Manages transition: heuristic_only → rl_assisted → rl_primary
  • Sleep cycle extraction task (Task 0) — Batch-processes unprocessed conversation logs to catch memories missed by real-time heuristics, with deduplication (similarity ≥ 0.90)
  • Sleep cycle log cleanup (Task 5) — Deletes processed logs older than 30 days (configurable)
  • Auto category classification — memory_save default category changed to "auto", which auto-classifies content using pattern matching
  • extraction_source metadata — Tracks how each memory was created: "heuristic", "rl", "auto", or "" (manual)
  • 6 new extraction config options in SleepCycleConfig: enable_memory_extraction, extraction_max_turns, extraction_dedup_threshold, extraction_min_info_density, extraction_rl_confidence_threshold, log_retention_days

Changed

  • SleepCycleRunner now accepts optional conversation_log parameter
  • SleepCycleReport includes extraction and log_cleanup_deleted fields
  • MemoryNode includes extraction_source field
  • GraphMemoryStore.add_memory() accepts extraction_source parameter

[0.3.0] - 2026-03-01

Changed

  • Project renamed from mcp-memory-server to long-term-memory. Python import name (aimemory) and CLI commands unchanged.
  • GitHub repository renamed to long-term-memory (old URL auto-redirects)

Added

  • Auto-linking — New memories automatically link to similar existing ones (similarity ≥ 0.92) via bidirectional graph edges
  • Live graph visualization (aimemory-live) — Real-time WebSocket-based memory graph in the browser
    • Force-directed layout (vis.js ForceAtlas2) with category-based coloring
    • Glow effects on save (green) and search (blue)
    • Event log sidebar with hover-to-highlight related nodes
    • Persistent event history across browser refreshes
    • Cross-process event push (MCP → live server via WebSocket /event path)
  • memory_visualize tool — Generate interactive static HTML graph (13th MCP tool)
  • [live] optional dependency — websockets>=12.0

Deprecated

  • PyPI package mcp-memory-server — use long-term-memory instead

[0.2.0] - 2026-02-28

First public release.

Added

  • MCP Server with 12 tools: auto_search, memory_save, memory_search, memory_update, memory_delete, memory_get_related, memory_pin, memory_unpin, memory_stats, sleep_cycle_run, policy_status, policy_decide
  • RL Memory Policy — Rule-based importance scoring + MLP contextual bandit (SAVE/SKIP/RETRIEVE)
  • Enhanced Policy (opt-in) — 778d state encoder (SentenceTransformer 768d + 10d hand-crafted), experience replay buffer, progressive autonomy
  • Semantic Search — ChromaDB vector store with intfloat/multilingual-e5-small embeddings
  • Knowledge Graph — NetworkX-based entity-relation graph with multi-hop traversal
  • GraphRAG Hybrid Retrieval (opt-in) — Vector similarity + graph traversal with RL re-ranker (11d features)
  • Multi-Resolution Text — 3 levels (full text, keyword summary, entity triples) with token-budget-aware composition
  • Forgetting Pipeline — Decay-based aging with compress → deactivate → delete stages
  • Sleep Cycle — Periodic maintenance: consolidation, resolution regeneration, forgetting, checkpoint saving
  • Immutable Memories — Protected from modification/deletion with SHA-256 rule hash verification
  • Pin/Unpin — User-controlled forgetting protection
  • Multilingual Support — Korean (ko) and English (en) i18n patterns
  • P2P Federated Learning — Gossip protocol with differential privacy and Krum aggregation
  • A/B Comparison Framework — Baseline vs re-ranked retrieval comparison
  • CI/CD — GitHub Actions with Python 3.11/3.12/3.13 matrix
  • 611 tests passing across 31 test files

Removed

  • Deprecated DualHeadDQN offline training pipeline (replaced by online MLP bandit)
  • ExtractorConfig (unused after DQN removal)