Mnema

What the mnema MCP server does

The mnema MCP server provides durable memory for AI agents through the Model Context Protocol. It connects MCP clients to a shared Mnema store where agents can save facts, retrieve relevant knowledge, inspect relationships, and browse recent entries.

By default, the store is local to the machine and lives under ~/.mnema. This makes the same memory available to supported clients, scripts, and CI processes that use the same Mnema instance. The repository distributes release bundles and binaries; the product documentation and downloads are hosted separately.

Mnema's recall combines lexical matching with graph spreading activation. Optional semantic embeddings are supported, and responses include a confidence band to indicate the certainty of a result.

How it works

Running mnema mcp starts the MCP interface. When configured for the default local endpoint, the command also starts the local store if it is not already running. MCP clients then receive four tools:

• consult searches memory before an action.
• learn records a durable fact.
• related returns neighboring items in the memory graph.
• list browses recent memory entries.

The underlying store uses an append-only segment format. The project describes its embedded engram database as formally specified and machine-checked, with property tests, CRDT conformance checking, Kani proofs, and a TLA+ model included in its release verification process.

Setup and configuration

To install the mnema MCP server, use an MCP Registry-aware client with the registry identifier ai.zimac/mnema, or download a platform .mcpb bundle from the project releases. The documented bundles target macOS universal and Linux x86_64. A standalone binary can also be installed with the published shell installer; unsupported platforms build from source.

A generic MCP configuration is:

{
  "mcpServers": {
    "mnema": {
      "command": "mnema",
      "args": ["mcp"]
    }
  }
}

The mnema MCP server reads these variables when configuration needs to differ from the defaults:

• MNEMA_DATA_DIR changes the local store directory. Its default is ~/.mnema.
• MNEMA_URL selects the store endpoint. Its default is http://127.0.0.1:7077.
• MNEMA_TOKEN supplies the bearer credential for a remote store.

Set MNEMA_URL and MNEMA_TOKEN to connect through an authenticated bridge to a remote mnema serve instance or hosted learnings pod. For a non-loopback URL, the local store is not auto-started.

Tools and capabilities

The four exposed tools cover the core memory workflow: retrieve knowledge with consult, persist new knowledge with learn, follow graph relationships with related, and inspect recent records with list. A single store can be used by Claude Code, Claude Desktop, Cursor, scripts, and CI, provided each connection targets the same local or remote Mnema instance.

Limitations and notes

The documented prebuilt downloads cover macOS universal and Linux x86_64. Other platforms use a source build through the installer rather than a listed prebuilt binary. The macOS GUI application is distributed separately from the headless binary.

Local operation does not require a cloud account, but remote operation requires an endpoint and bearer token. The README does not specify retention controls, backup commands, embedding-provider configuration, or the exact behavior of the confidence band, so those details should not be assumed from the MCP interface alone.