Memory MCP Server (Embedded GraphRAG + Vector Memory)

Persistent knowledge graph memory and context window caching for LLM conversations. Built on an embedded KuzuDB Cypher graph database with local HuggingFace vector embeddings (Xenova/all-MiniLM-L6-v2) — zero external database setup required.

An MCP (Model Context Protocol) server that provides hybrid semantic vector + multi-hop Cypher graph memory retrieval and context window caching for AI coding environments like Cursor and Claude Desktop.

Why Use Memory MCP?

• Embedded Knowledge Graph (KuzuDB) – Stores entities, observations, and typed relationships in a local, in-process Cypher graph database (~/.memory-mcp/graph_db) with zero external setup.
• Hybrid Vector + Graph Retrieval – Uses local 384-d embeddings (Xenova/all-MiniLM-L6-v2) and Kuzu array_cosine_similarity to find entry nodes, then traverses 1–3 relationship hops via Cypher so the LLM only receives compact, high-signal facts.
• Context Window Caching – Intelligent archiving, semantic relevance scoring, and graph-linked summaries (Summarizes / Mentions edges).
• Zero-Config Install – No MongoDB or Neo4j server required.

Quick Start

Install in Cursor (Recommended):

Install in Cursor

Or install manually:

npm install -g @jamesanz/memory-mcp
# Or from source:
git clone https://github.com/JamesANZ/memory-mcp.git
cd memory-mcp && npm install && npm run build

Features

Knowledge Graph & Hybrid Retrieval Tools

• store-graph-memory – Upsert structured entities (name, entityType, observations[]) and directed relations (from, to, relationType) with automatic 384-d embeddings.
• query-memory – Hybrid semantic vector + multi-hop Cypher traversal query that returns only the relevant connected subgraph and memory chunks.
• remove-graph-entities – Remove outdated entities or specific relationships so contradictory facts do not linger.

Basic Memory Tools

• save-memories – Save memories to the graph database (overwrites existing)
• get-memories – Retrieve all stored memory logs
• add-memories – Append new memories and auto-extract entity mentions
• clear-memories – Remove all stored memories and graph entities

Context Window Caching

• archive-context – Archive conversation context with tags and vector embeddings
• retrieve-context – Retrieve relevant archived context
• score-relevance – Score archived content relevance using vector cosine similarity
• create-summary – Create summaries of archived content linked via Summarizes edges
• get-conversation-summaries – Get all summaries for a conversation
• search-context-by-tags – Search archived content by tags

Configuration

By default, the graph database and local embedding model cache are stored in ~/.memory-mcp/. You can customize these paths via environment variables:

export MEMORY_DB_PATH="/custom/path/to/graph_db"
export MEMORY_MODEL_CACHE_DIR="/custom/path/to/models"

Claude Desktop

Add to claude_desktop_config.json:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "memory-mcp": {
      "command": "node",
      "args": ["/absolute/path/to/memory-mcp/build/index.js"]
    }
  }
}

Usage Examples

Store Structured Graph Memory

{
  "tool": "store-graph-memory",
  "arguments": {
    "entities": [
      {
        "name": "memory-mcp",
        "entityType": "Project",
        "observations": ["Uses embedded KuzuDB graph database", "Written in TypeScript with ESM"]
      },
      {
        "name": "KuzuDB",
        "entityType": "Tech",
        "observations": ["In-process property graph database supporting Cypher and vector similarity"]
      }
    ],
    "relations": [
      {
        "from": "memory-mcp",
        "to": "KuzuDB",
        "relationType": "USES_DATABASE"
      }
    ],
    "llm": "claude"
  }
}

Query Subgraph Memory

{
  "tool": "query-memory",
  "arguments": {
    "query": "What database architecture does memory-mcp use?",
    "maxHops": 2,
    "limit": 5
  }
}

Technical Details

Built with: Node.js, TypeScript, MCP SDK, KuzuDB (kuzu), HuggingFace Transformers (@huggingface/transformers), zod
Storage: Embedded KuzuDB (~/.memory-mcp/graph_db)

License

MIT