mcp-memory-graph

A context-aware memory MCP server for Claude Code and any MCP-compatible AI agent.

Goes beyond basic vector search by adding authority weighting, conflict detection, and typed relationship edges between memories — so your agent always retrieves the right answer when sources disagree.

Inspired by the context engine architecture described in Unblocked's "How a Context Engine Actually Works".

Why this exists

Standard memory MCP servers store and retrieve memories by semantic similarity. That works until you have conflicting memories — an old instruction saying one thing and a new one saying another. Without authority weighting, the agent retrieves whichever is semantically closer to the query, not whichever is more trustworthy.

mcp-memory-graph solves this with three mechanisms:

Installation

pip install mcp-memory-graph

Or run directly:

git clone https://github.com/RetroRobAI/mcp-memory-graph
cd mcp-memory-graph
pip install -r requirements.txt
python server.py

Claude Code setup

Add to ~/.claude.json under mcpServers:

"mcp-memory-graph": {
  "type": "stdio",
  "command": "mcp-memory-graph",
  "env": {
    "MEMORY_GRAPH_DB_PATH": "/path/to/memories.db"
  }
}

Or with the raw script:

"mcp-memory-graph": {
  "type": "stdio",
  "command": "python",
  "args": ["/path/to/mcp-memory-graph/server.py"],
  "env": {
    "MEMORY_GRAPH_DB_PATH": "/path/to/memories.db"
  }
}

Migrating from an existing memory service

If you have an existing memory service (mcp-memory-service, Mem0, or a markdown-based memory system), you can import your memories into mcp-memory-graph using the included migration script.

Migration is manual and opt-in — it never runs automatically. Nothing is written until you explicitly confirm.

Run the migration script

python -m mcp_memory_graph.migrate

The script will:

1. Auto-detect any existing mcp-memory-service SQLite database
2. Ask if you have a markdown memory directory to import
3. Show you how many memories it found
4. Present three choices:
   • [1] Migrate — import everything into mcp-memory-graph
   • [2] Run in parallel — start mcp-memory-graph fresh, keep your old service running
   • [3] Skip — do nothing
5. Ask for a final confirmation before writing anything

Your existing memory service is never modified — the script only reads from it.

Configuration

All settings via environment variables:

Tools

... View full README on GitHub