A-MEM MCP Server

Brain-like memory for Claude Code based on the NeurIPS 2025 paper "A-MEM: Agentic Memory for LLM Agents".

Features

Automatic memory linking: New memories automatically connect to related existing ones
Memory evolution: New information updates the context of existing memories
Rich metadata: Auto-generated keywords, context, and tags
Global storage: Shared across all projects at ~/.codeagent/memory/
Fallback mode: Works without A-MEM library using simple keyword matching

Tools

| Tool | Description | |------|-------------| | store_memory | Store knowledge with automatic linking and evolution | | search_memory | Semantic search across all memories | | read_memory | Read specific memory by ID with full metadata | | list_memories | List recent memories with filtering | | update_memory | Update existing memory (triggers re-evolution) | | delete_memory | Remove a memory | | get_memory_stats | Statistics about the memory system |

Philosophy

Based on A-MEM (2024) - "Agentic Memory for LLM Agents".

Key insight: Memory should behave like a brain, not a database. New memories automatically link to related ones (Zettelkasten-style), and existing memories evolve when new information arrives.

Installation

Basic (fallback mode - JSON storage)

pip install git+https://github.com/Questi0nM4rk/amem-mcp.git

Full (ChromaDB + semantic search)

pip install "git+https://github.com/Questi0nM4rk/amem-mcp.git[full]"

Note: Full installation requires:

OPENAI_API_KEY environment variable for metadata generation
OR Ollama running locally

Usage with Claude Code

claude mcp add amem -- python -m amem_mcp.server

Usage

# Store a memory
store_memory(content="JaCore uses repository pattern with Unit of Work")

# Search memories
search_memory(query="data access patterns")

# Read specific memory
read_memory(memory_id="mem_0001")

How It Works

Store: Content is analyzed, keywords extracted, and similar memories found
Link: Bidirectional links created between related memories
Evolve: Existing memories' context updated with new information
Search: Vector similarity + link traversal for comprehensive results

Backend

| Feature | Full Mode | Fallback Mode | |---------|-----------|---------------| | Storage | ChromaDB | JSON file | | Search | Semantic vectors | Keyword matching | | Dependencies | chromadb, sentence-transformers | None | | Accuracy | High | Lower |

Full mode: Uses ChromaDB for vector storage with semantic search
Fallback mode: JSON file with keyword-based matching (still functional!)

Environment Variables

| Variable | Required | Description | |----------|----------|-------------| | OPENAI_API_KEY | For full mode | LLM for metadata generation | | CODEAGENT_HOME | Optional | Override default ~/.codeagent |

Storage

Memories stored at: ~/.codeagent/memory/

ChromaDB (full mode): ~/.codeagent/memory/chromadb/
JSON (fallback): ~/.codeagent/memory/memories.json

MCP Servers