memcp

Backend-agnostic, multi-tenant MCP memory server. AI clients connect and get persistent long-term memory over streamable HTTP.

Quickstart

Start memcp on 127.0.0.1:8080 with the sqlite backend, using Docker or pip. For search by meaning and fact extraction, use mem0 instead.

With Docker

Needs Docker. The command names your key and its space after your login, $USER. A key name holds only letters, digits, _, . and -: if your login has anything else, or USER is unset, replace $USER with such a name.

mkdir memcp && cd memcp &&
  curl -fsSLO https://raw.githubusercontent.com/Jartan-LLC/memcp/main/docker-compose.yml &&
  curl -fsSL -o .env https://raw.githubusercontent.com/Jartan-LLC/memcp/main/.env.example &&
  chmod 600 .env &&
  TOKEN=$(openssl rand -hex 32) && echo "$TOKEN" &&
  printf '[[keys]]\nname = "%s"\nsecrets = ["sha256:%s"]\ngrants = [{ space = "%s", relation = "writer" }]\n' \
    "$USER" "$(printf %s "$TOKEN" | sha256sum | cut -d' ' -f1)" "$USER" > keys.toml

Keep the token that printed; keys.toml holds only its hash (on macOS, use shasum -a 256 in place of sha256sum). Then start it:

docker compose up -d --wait

With pip

Needs Python 3.12 or later.

pip install memcp-server
MEMCP_BACKEND=sqlite MEMCP_SQLITE_PATH="$HOME/.memcp/memcp.sqlite3" MEMCP_HOST=127.0.0.1 memcp

Where the system Python refuses pip install (an "externally managed" environment), use pipx install memcp-server. With no MEMCP_KEYS_FILE, memcp serves one key on loopback and clients connect without a token.

Connect from Claude Code

With the token you generated:

claude mcp add --scope user --transport http memcp http://localhost:8080/mcp \
  --header "Authorization: Bearer YOUR_TOKEN"

Behind a reverse proxy, the URL is https://<your-host>/mcp. A server started with memcp on loopback and no MEMCP_KEYS_FILE takes no --header.

Try it

Ask Claude to remember something:

"Remember that I prefer Python 3.12 and use ruff for linting."

In a new conversation, ask:

"What linter do I use?"

Claude searches memory automatically and uses the stored context. On the quickstart's sqlite backend this works because linter and linting share a stem — matching is lexical, so a question phrased in words that do not appear in the memory will not find it. mem0 is the backend that matches on meaning.

What it does

• Durable, keyless local storage (sqlite), or mem0 on pgvector for semantic search and a knowledge graph
• Add, search, list, update and delete memories over MCP — 12 tools, frozen surface
• Flat scope-based filtering (agent_id, run_id)
• Bearer-token keys, each read-only or read-write on one space, stored as hashes
• Stateless HTTP transport — safe behind reverse proxies

Documentation

Status

v0.2.0 — API may change before v1.0. Suitable for development and early adoption.

License

AGPL-3.0 — see LICENSE.