MCP AI Memory

A production-ready Model Context Protocol (MCP) server for semantic memory management that enables AI agents to store, retrieve, and manage contextual knowledge across sessions.

📖 System Prompt Available: See SYSTEM_PROMPT.md for a comprehensive guide on how to instruct AI models to use this memory system effectively. This prompt helps models understand when and how to use memory tools, especially for proactive memory retrieval.

Features

Core Capabilities

• TypeScript - Full type safety with strict mode

• PostgreSQL + pgvector - Vector similarity search with HNSW indexing

• Kysely ORM - Type-safe SQL queries

• Local Embeddings - Uses Transformers.js (no API calls)

• Intelligent Caching - Redis + in-memory fallback for blazing fast performance

• Multi-Agent Support - User context isolation

• Token Efficient - Embeddings removed from responses

Advanced Memory Management

• Graph Relationships - Rich relationship types (references, contradicts, supports, extends, causes, precedes, etc.)

• Graph Traversal - BFS/DFS algorithms with depth limits and filtering

• Memory Decay - Automatic lifecycle management with exponential decay

• Memory States - Active, dormant, archived, and expired states

• Preservation - Protect important memories from decay

• Soft Deletes - Data recovery with deleted_at timestamps

• Clustering - Automatic memory consolidation

• Compression - Automatic compression of archived memories

Prerequisites

• Node.js 18+ or Bun

• PostgreSQL with pgvector extension

• Redis (optional - falls back to in-memory cache if not available)

Installation

NPM Package (Recommended for Claude Desktop)

From Source

1. Install dependencies:

2. Set up PostgreSQL with pgvector:

3. Create environment file:

4. Run migrations:

Usage

Development

Production

Troubleshooting

Embedding Dimension Mismatch Error

If you see an error like:

This occurs when the embedding model changes between sessions. To fix:

1. Option 1: Reset and Re-embed (Recommended for new installations)

   # Clear existing memories and start fresh psql -d your_database -c "TRUNCATE TABLE memories CASCADE;"

2. Option 2: Specify a Consistent Model Add EMBEDDING_MODEL to your Claude Desktop config:

   { "mcpServers": { "memory": { "command": "npx", "args": ["-y", "mcp-ai-memory"], "env": { "MEMORY_DB_URL": "postgresql://...", "EMBEDDING_MODEL": "Xenova/all-mpnet-base-v2" } } } }

   Common models:

   • Xenova/all-mpnet-base-v2 (768 dimensions - default, best quality)

   • Xenova/all-MiniLM-L6-v2 (384 dimensions - smaller/faster)

3. Option 3: Run Migration for Flexible Dimensions If you're using the source version:

   bun run migrate

   This allows mixing different embedding dimensions in the same database.

Database Connection Issues

Ensure your PostgreSQL has the pgvector extension:

Claude Desktop Integration

💡 For Best Results: Include the SYSTEM_PROMPT.md content in your Claude Desktop system prompt or initial conversation to help Claude understand how to use the memory tools effectively.

Quick Setup (NPM)

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

With Optional Redis Cache

Environment Variables

Available Tools

💡 Token Efficiency: Default limits are set to 10 results to optimize token usage. Increase only when needed.

Core Operations (Most Important)

• memory_search - SEARCH FIND RECALL - Search stored information using natural language (USE THIS FIRST! Default limit: 10)

• memory_list - LIST BROWSE SHOW - List all memories chronologically (fallback when search fails, default limit: 10)

• memory_store - STORE SAVE REMEMBER - Store new information after checking for duplicates

• memory_update - UPDATE MODIFY EDIT - Update existing memory metadata

• memory_delete - DELETE REMOVE FORGET - Delete specific memories

Advanced Operations

• memory_batch - BATCH BULK IMPORT - Store multiple memories efficiently

• memory_batch_delete - Delete multiple memories at once

• memory_graph_search - GRAPH RELATED - Search with relationship traversal (alias for memory_traverse)

• memory_consolidate - MERGE CLUSTER - Group similar memories

• memory_stats - STATS INFO - Database statistics

• memory_relate - LINK CONNECT - Create memory relationships

• memory_unrelate - UNLINK DISCONNECT - Remove relationships

• memory_get_relations - Show all relationships for a memory

Graph & Decay Operations (New)

• memory_traverse - TRAVERSE EXPLORE - Traverse memory graph with BFS/DFS algorithms

• memory_graph_analysis - ANALYZE CONNECTIONS - Analyze graph connectivity and relationship patterns

• memory_decay_status - DECAY STATUS - Check decay status of a memory

• memory_preserve - PRESERVE PROTECT - Preserve important memories from decay

Resources

• memory://stats - Database statistics

• memory://types - Available memory types

• memory://tags - All unique tags

• memory://relationships - Memory relationships

• memory://clusters - Memory clusters

Prompts

• load-context - Load relevant context for a task

• memory-summary - Generate topic summaries

• conversation-context - Load conversation history

Architecture

Environment Variables

Caching Architecture

The server implements a two-tier caching strategy:

1. Redis Cache (if available) - Distributed, persistent caching

2. In-Memory Cache (fallback) - Local NodeCache for when Redis is unavailable

Async Job Processing

When Redis is available and ENABLE_ASYNC_PROCESSING=true, the server uses BullMQ for background job processing:

Features

• Async Embedding Generation: Offloads CPU-intensive embedding generation to background workers

• Batch Import: Processes large memory imports without blocking the main server

• Memory Consolidation: Runs clustering and merging operations in the background

• Automatic Retries: Failed jobs are retried with exponential backoff

• Dead Letter Queue: Permanently failed jobs are tracked for manual intervention

Running Workers

Queue Monitoring

The memory_stats tool includes queue statistics when async processing is enabled:

• Active, waiting, completed, and failed job counts

• Processing rates and performance metrics

• Worker health status

Cache Invalidation

• Memory updates/deletes automatically invalidate relevant caches

• Search results are cached with query+filter combinations

• Embeddings are cached for 24 hours (configurable)

Development

Type Checking

Linting

Using with AI Models

System Prompt for Better Memory Usage

The memory tools include enhanced descriptions with keywords to help models understand when to use each tool. However, for best results with models like Gemma3, Qwen, or other open-source models:

1. Include the System Prompt: Copy the content from SYSTEM_PROMPT.md and include it in your initial conversation or system prompt

2. Key Behaviors to Reinforce:

   • Always use memory_search FIRST before any operation

   • Use memory_list as a fallback when search returns no results

   • Search for user information at conversation start (e.g., "user name preferences")

   • Store structured JSON in the content field

Example Initial Prompt for Models

Implementation Status

✅ Fully Integrated Features

• DBSCAN Clustering: Advanced clustering algorithm for memory consolidation

• Smart Compression: Automatic compression for large memories (>100KB)

• Context Window Management: Token counting and intelligent truncation

• Input Sanitization: Comprehensive validation and sanitization

• All Workers Active: Embedding, batch, and clustering workers all operational

Testing

The project includes a comprehensive test suite covering:

• Memory service operations (store, search, update, delete)

• Input validation and sanitization

• Clustering and consolidation

• Compression for large content

Run tests with bun test.

License

MIT