mem0-mcp

A small MCP server that wraps a self-hosted mem0 REST API (see ../server-compose-files/ai-containers) as tools for Claude Desktop / Claude Code. Mem0's own official MCP integrations (mcp.mem0.ai, the archived mem0-mcp-server package) only target Mem0's cloud API — different auth scheme (Authorization: Token ... vs. this server's X-API-Key) and different routes — so they don't work against a self-hosted instance. This wraps the self-hosted REST API directly instead.

Tools

Tool	mem0 endpoint
add_memory	POST /memories
search_memories	POST /search
list_memories	GET /memories
get_memory	GET /memories/{id}
update_memory	PUT /memories/{id}
delete_memory	DELETE /memories/{id}

Bulk operations (delete_all_memories, /reset) are intentionally not exposed — keeps the blast radius of an LLM-driven tool call small. Use curl directly against the API for those.

Setup

Requires an API key from the mem0 dashboard (/setup on first run, or the API Keys page after) and the URL your machine can reach the mem0 API at.

Add to Claude Desktop's config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS). Two ways to run it, depending on whether you've cloned the repo:

Without cloning, straight from GitHub — same uvx --from git+... pattern other MCP servers use to run from a repo without publishing to PyPI first:

uvx --from git+https://github.com/andresmorales07/mem0-mcp mem0-mcp

{
  "mcpServers": {
    "mem0": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/andresmorales07/mem0-mcp", "mem0-mcp"],
      "env": {
        "MEM0_API_URL": "http://your-mem0-host:8888",
        "MEM0_API_KEY": "m0sk_...",
        "MEM0_DEFAULT_USER_ID": "your-user-id",
        "MEM0_DEFAULT_AGENT_ID": "claude-desktop"
      }
    }
  }
}

Prefer to pin to a released version instead of tracking main? Append a ref, e.g. git+https://github.com/andresmorales07/mem0-mcp@v0.1.0.

From a local clone:

{
  "mcpServers": {
    "mem0": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/mem0-mcp", "mem0-mcp"],
      "env": {
        "MEM0_API_URL": "http://your-mem0-host:8888",
        "MEM0_API_KEY": "m0sk_...",
        "MEM0_DEFAULT_USER_ID": "your-user-id",
        "MEM0_DEFAULT_AGENT_ID": "claude-desktop"
      }
    }
  }
}

Use the full path to uv/uvx (e.g. which uv) if Claude Desktop can't find them on its own PATH.

Both default env vars are optional, and they behave differently:

• MEM0_DEFAULT_USER_ID is a last-resort fallback — only used when a tool call supplies no user_id, agent_id, or run_id at all. Applies to both reads and writes.

• MEM0_DEFAULT_AGENT_ID is merged in only on add_memory (unless that call passes its own agent_id), tagging every memory this server writes with which client created it — useful if you ever point another tool (n8n, a different agent) at the same mem0 instance and want to tell them apart later.

  It's deliberately not applied to search_memories/list_memories. MCP tool calls can't distinguish "no agent_id given" from "search across all agents," so auto-tagging reads would silently and permanently scope every query to just this one agent — making it impossible to ever ask "show me everything anyone has learned about me" again. Reads default to whatever user_id/agent_id/run_id the model explicitly passes (plus the MEM0_DEFAULT_USER_ID fallback above); pass agent_id explicitly on a read if you want to narrow it to one agent's memories.

Restart Claude Desktop after editing the config. Claude will prompt for permission the first time it calls one of these tools.

Manual test

MEM0_API_URL=http://your-mem0-host:8888 MEM0_API_KEY=m0sk_... uv run python -c "
from mem0_mcp.server import list_memories
print(list_memories(user_id='alice'))
"