02-long-term-memory

AgentCore Memory: Long-Term Memory Strategies

Overview

Long-term memory in Amazon Bedrock AgentCore enables AI agents to maintain persistent information across multiple conversations and sessions. Unlike short-term memory that focuses on immediate context, long-term memory extracts, processes, and stores meaningful information that can be retrieved and applied in future interactions, creating truly personalized and intelligent agent experiences.

What is Long-Term Memory?

Long-term memory provides:

• Cross-Session Persistence: Information that survives beyond individual conversations
• Intelligent Extraction: Automatic identification and storage of important facts, preferences, and patterns
• Semantic Understanding: Vector-based storage that enables natural language retrieval
• Personalization: User-specific information that enables tailored experiences
• Knowledge Accumulation: Continuous learning and information building over time

How Long-Term Memory Strategies Work

Long-term memory operates through Memory Strategies that define what information to extract and how to process it. The system works automatically in the background:

Processing Pipeline

1. Conversation Analysis: Saved conversations are analyzed based on configured strategies
2. Information Extraction: Important data (facts, preferences, summaries) is extracted using AI models
3. Structured Storage: Extracted information is organized in namespaces for efficient retrieval
4. Semantic Indexing: Information is vectorized for natural language search capabilities
5. Consolidation: Similar information is merged and refined over time

Processing Time: Typically takes ~1 minute after conversations are saved, with no additional code required.

Behind the Scenes

• AI-Powered Extraction: Uses foundation models to understand and extract relevant information
• Vector Embeddings: Creates semantic representations for similarity-based retrieval
• Namespace Organization: Structures information using configurable path-like hierarchies
• Automatic Consolidation: Merges and refines similar information to prevent duplication
• Incremental Learning: Continuously improves extraction quality based on conversation patterns

Long-Term Memory Strategy Types

AgentCore Memory supports four distinct strategy types for long-term information storage:

1. Semantic Memory Strategy

Stores factual information extracted from conversations using vector embeddings for similarity search.

{
    "semanticMemoryStrategy": {
        "name": "FactExtractor",
        "description": "Extracts and stores factual information",
        "namespaces": ["support/user/{actorId}/facts/"]
    }
}

Best for: Storing product information, technical details, or any factual data that needs to be retrieved through natural language queries.

2. Summary Memory Strategy

Creates and maintains summaries of conversations to preserve context for long interactions.

{
    "summaryMemoryStrategy": {
        "name": "ConversationSummary",
        "description": "Maintains conversation summaries",
        "namespaces": ["support/summaries/{sessionId}/"]
    }
}

Best for: Providing context in follow-up conversations and maintaining continuity across long interactions.

3. User Preference Memory Strategy

Tracks user-specific preferences and settings to personalize interactions.

{
    "userPreferenceMemoryStrategy": {
        "name": "UserPreferences",
        "description": "Captures user preferences and settings",
        "namespaces": ["support/user/{actorId}/preferences"/]
    }
}

Best for: Storing communication preferences, product preferences, or any user-specific settings.

4. Custom Memory Strategy

Allows customization of prompts for extraction and consolidation, providing flexibility for specialized use cases.

{
    "customMemoryStrategy": {
        "name": "CustomExtractor",
        "description": "Custom memory extraction logic",
        "namespaces": ["user/custom/{actorId}/"],
        "configuration": {
            "semanticOverride": { # You can also override Summary or User Preferences.
                "extraction": {
                    "appendToPrompt": "Extract specific information based on custom criteria",
                    "modelId": "global.anthropic.claude-haiku-4-5-20251001-v1:0",
                },
                "consolidation": {
                    "appendToPrompt": "Consolidate extracted information in a specific format",
                    "modelId": "global.anthropic.claude-haiku-4-5-20251001-v1:0",
                }
            }
        }
    }
}

Best for: Specialized extraction needs that don't fit the standard strategies.

Understanding Namespaces

Namespaces organize memory records within strategies using a path-like structure. They can include variables that are dynamically replaced:

• support/facts/{sessionId}: Organizes facts by session
• user/{actorId}/preferences: Stores user preferences by actor ID
• meetings/{memoryId}/summaries/{sessionId}: Groups summaries by memory

The {actorId}, {sessionId}, and {memoryId} variables are automatically replaced with actual values when storing and retrieving memories.

Example: How It Works in Practice

Let's say a user tells your customer support agent: "I'm vegetarian and I really enjoy Italian cuisine. Please don't call me after 6 PM."

After you save this conversation, the configured strategies automatically:

Semantic Strategy extracts:

• "User is vegetarian"
• "User enjoys Italian cuisine"

User Preference Strategy captures:

• "Dietary preference: vegetarian"
• "Cuisine preference: Italian"
• "Contact preference: no calls after 6 PM"

Summary Strategy creates:

• "User discussed dietary restrictions and contact preferences"

All of this happens automatically in the background - you only need to save the conversation and the strategies handle the rest.

Available Sample Notebooks

Explore these hands-on examples to learn long-term memory strategy implementation:

Integration Method	Use Case	Description	Notebook	Architecture
Strands Agent Hooks	Customer Support	Complete support system with semantic and preference memory strategies	customer-support.ipynb	View
Strands Agent Hooks	Math Assistant	Math tutor assistant that remembers user learning preferences and progress	math-assistant.ipynb	View
LangGraph Agent Hooks	Nutrition Assistant	Nutrition advisor that saves user dietary preferences and health goals for personalized recommendations	nutrition-assistant-with-user-preference-saving.ipynb	View
Strands Agent Memory Tool	Culinary Assistant	Food recommendation agent that learns dietary preferences and cooking styles	culinary-assistant.ipynb	View
Multi-Agent	Agent Collaboration	Travel Assistant with multiple agents sharing and utilizing long-term memory strategies	travel-booking-assistant.ipynb	View

Getting Started

1. Choose a sample that matches your use case
2. Navigate to the sample folder
3. Install requirements: pip install -r requirements.txt
4. Open the Jupyter notebook and follow the step-by-step implementation

Best Practices

1. Strategy Selection: Choose appropriate strategies based on your use case requirements
2. Namespace Design: Plan namespace hierarchies for efficient information organization
3. Extraction Tuning: Customize extraction prompts for domain-specific information
4. Performance Monitoring: Track memory extraction quality and retrieval performance
5. Privacy Considerations: Implement appropriate data retention and privacy policies

Next Steps

After mastering long-term memory strategies, explore:

• Combining short-term and long-term memory for comprehensive agent experiences
• Advanced custom strategy configurations
• Multi-agent memory sharing patterns
• Production deployment considerations