Memory MCP

Persistent memory and full-text session search for AI coding assistants, exposed as an MCP server.

The problem

AI coding assistants forget everything between sessions. Architecture decisions, user preferences, project context, what you debugged last Tuesday -- gone. You re-explain the same things constantly.

Memory MCP fixes this with two capabilities:

1. Explicit memory -- save notes, decisions, patterns, and preferences that persist across sessions. Your assistant remembers what you told it.

2. Session search -- full-text search across your entire conversation history. Find that thing you discussed three weeks ago without scrolling through logs.

No database servers. No background processes. No cloud. One SQLite file on your machine.

Related MCP server: MCP Vector Memory

Supported session sources

Adding a new source requires one parser file and a registry entry. See Adding a new source.

Installation

Requires Python 3.11+ with SQLite FTS5 support (included in standard Python builds).

pip install -e .

Or run directly with uv (no install needed):

uv run --directory /path/to/memory_mcp python -m memory_mcp

MCP configuration

Add to your MCP client config (e.g., ~/.claude/mcp.json or project-level .mcp.json):

With pip install:

{
  "mcpServers": {
    "memory": {
      "command": "memory-mcp"
    }
  }
}

With uv (no install):

{
  "mcpServers": {
    "memory": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/memory_mcp", "python", "-m", "memory_mcp"]
    }
  }
}

Tools

Memory (explicit knowledge store)

Sessions (historical conversation search)

How it works

On startup, Memory MCP yields its tool list to the MCP client immediately (<500 ms cold) and runs the initial session scan in a background task. The embedding model loads lazily on the first semantic search call — keyword search and saved memories work without it. Subsequent startups skip files whose mtime hasn't changed.

• Database location: ~/.memory_mcp/memory.db (override with MEMORY_MCP_DB env var)

• Session sources: auto-detected from standard locations (extend with MEMORY_MCP_SOURCES env var, format: type:path;type:path)

• Indexing: incremental by file mtime, parallelized across 8 threads

• Search: FTS5 with BM25 ranking, prefix matching, phrase support; optional vector search via sqlite-vec + fastembed (BAAI/bge-small-en-v1.5) when semantic=true is passed

• Startup: non-blocking — heavy work (scan, model load, vector backfill) runs after the server is already responding to tool calls

Adding a new session source

1. Create memory_mcp/parsers/your_source.py implementing the SessionParser protocol:

   • source_type: str attribute

   • parse_file(path: str) -> ParsedSession | None method

2. Register it in memory_mcp/parsers/__init__.py

3. Add directory detection in memory_mcp/config.py

See parsers/claude_code.py or parsers/omp.py for examples.

Testing

python tests/test_e2e.py        # end-to-end: spawns server, exercises all 9 tools
python tests/test_startup.py    # startup contract: cold Popen -> tools/list under 1.5s

test_e2e.py starts the MCP server as a subprocess, exercises all 9 tools over the stdio protocol, and asserts tool responses. test_startup.py enforces the v0.3.0 startup contract — if an eager import or pre-yield blocking call regresses startup speed, it fails immediately. Both use throwaway databases so your real data is untouched.

Architecture

memory_mcp/
  server.py        # FastMCP entry point, lifespan yields fast then runs scan in background
  readiness.py     # Lazy embedder + scan/backfill coordination
  config.py        # Auto-detects session dirs, DB path
  db.py            # SQLite + FTS5 + sqlite-vec schema, all queries, sync triggers
  embeddings.py    # Lazy fastembed wrapper (BAAI/bge-small-en-v1.5)
  scanner.py       # Walks session dirs, dispatches to parsers, parallel indexing
  parsers/
    base.py        # ParsedSession / ParsedMessage dataclasses, SessionParser protocol
    claude_code.py # Claude Code JSONL parser (merges streamed assistant blocks)
    claude_history.py # Claude Code history.jsonl parser (one file, many sessions)
    omp.py         # OMP JSONL parser
    opencode.py    # OpenCode SQLite parser (reads DB directly, read-only)
  tools/
    memory.py      # save_memory, search_memory, list_memories, delete_memory
    sessions.py    # list_sessions, get_session, search_sessions, refresh_sessions

License

MIT