memory-mcp

Persistent, searchable, versioned memory for AI agents — backed by Valkey (Redis-compatible), exposed as an MCP server over HTTP.

Works with any MCP-compatible agent: Claude Code, Cursor, VS Code, and others.

What it does

• Stores named memory entries with tags, types, and project scopes

• Tag-intersection search, type/project filtering, and substring search

• Hit tracking (entries accessed more float to the top)

• Full version history with rollback

• Prometheus metrics endpoint

• Optional bearer token auth

Quick start

cp .env.example .env
# Optional: set MEMORY_MCP_AUTH_TOKEN in .env (see Auth section)
docker compose up -d

This pulls the pre-built image from GHCR. The MCP server is now available at http://127.0.0.1:3106/mcp.

To build locally instead:

docker compose build
docker compose up -d

Using an existing Redis or Valkey

By default docker compose up -d starts a bundled Valkey container. To connect to an existing Redis or Valkey instance instead, set VALKEY_URL and start only the memory-mcp service:

# .env
VALKEY_URL=redis://your-host:6379

docker compose up -d memory-mcp

Any Redis-compatible server (Redis 6+, Valkey, KeyDB, Upstash via rediss://, etc.) works. The server uses only basic data structures: hashes, lists, and sets.

Agent setup

Copy AGENTS.md from this repo into your project root. It tells your agent how to use the memory tools, what to store, and when.

Then register the MCP server with your agent client:

Claude Code

# Without auth
claude mcp add memory --transport http http://127.0.0.1:3106/mcp

# With auth
claude mcp add memory --transport http http://127.0.0.1:3106/mcp \
  --header "Authorization: Bearer your-token"

Or add manually to ~/.claude.json:

{
  "mcpServers": {
    "memory": {
      "type": "http",
      "url": "http://127.0.0.1:3106/mcp",
      "headers": { "Authorization": "Bearer your-token" }
    }
  }
}

Cursor

Add to ~/.cursor/mcp.json (global) or .cursor/mcp.json (project):

{
  "mcpServers": {
    "memory": {
      "url": "http://127.0.0.1:3106/mcp",
      "headers": { "Authorization": "Bearer your-token" }
    }
  }
}

VS Code (GitHub Copilot, MCP extension)

Add to .vscode/mcp.json in your project:

{
  "servers": {
    "memory": {
      "type": "http",
      "url": "http://127.0.0.1:3106/mcp",
      "headers": { "Authorization": "Bearer your-token" }
    }
  }
}

Omit the headers / Authorization line in any config if you are not using auth.

Configuration

Copy .env.example to .env and edit as needed.

Auth

By default the server runs unauthenticated. This is safe when bound to loopback (127.0.0.1) and accessed only from the local machine.

To enable auth:

# Generate a token
openssl rand -hex 32

# Add to .env
MEMORY_MCP_AUTH_TOKEN=your-generated-token

docker compose up -d

All requests to POST /mcp must then include:

Authorization: Bearer <token>

GET /health and GET /metrics are always unauthenticated.

Available tools

Memory types

pattern, decision, reference, feedback, incident, project, entity, state

Endpoints

Data model

Each entry is stored as a Redis hash at mem:<id>:

Version history is stored in a Redis list at memver:<id> (newest-first, capped at MAX_VERSIONS_PER_ENTRY).

Tag, type, and project indexes are Redis sets (tag:<name>, type:<name>, project:<name>).

License

MIT