mcp-memory-server

A persistent semantic memory layer for AI coding agents, built with FastAPI and the Model Context Protocol (MCP).

Agents connect over MCP and use two tools -- store_memory to save text with vector embeddings, and retrieve_memory to search by meaning. Weaviate handles vector storage; sentence-transformers generates embeddings client-side.

Architecture

MCP Client (agent)
    |
    |  streamable HTTP  (/mcp)
    v
FastAPI + FastMCP
    |
    |-- sentence-transformers  (all-MiniLM-L6-v2)
    |-- Weaviate               (local, no vectorizer modules)

Related MCP server: umo-memory

Prerequisites

• Python 3.11+

• Docker (for Weaviate)

• gh CLI (optional, for creating the GitHub repo)

Quickstart

# 1. Start Weaviate
docker compose up -d

# 2. Create a virtual environment and install dependencies
python -m venv .venv
.venv\Scripts\activate      # Windows
# source .venv/bin/activate  # macOS / Linux
pip install -r requirements.txt

# 3. Run the server
python -m memory_server

The server starts on http://localhost:8000.

MCP Tools

store_memory

Save text with an auto-generated embedding.

Returns { "id": "<uuid>", "status": "stored" }.

retrieve_memory

Semantic search over stored memories.

Returns a list of matches with text, metadata fields, and a distance score (lower = more relevant).

Project Structure

mcp-memory-server/
  memory_server/
    __init__.py
    __main__.py       # python -m entrypoint
    main.py           # FastAPI app + MCP tool definitions
    store.py          # Weaviate client + embedding logic
  docker-compose.yml  # Local Weaviate instance
  requirements.txt

License

MIT