Governed MCP memory server

Configure

Add Heartwood to your MCP client config. Override HEARTWOOD_DB_PATH to share the same store built by import-markdown or the warm recall service.

{
  "mcpServers": {
    "heartwood-memory": {
      "command": "python",
      "args": ["-m", "heartwood.adapters.mcp_server"],
      "env": {
        "HEARTWOOD_DB_PATH": ".heartwood/heartwood.db",
        "HEARTWOOD_TENANT": "tenant:ops"
      }
    }
  }
}

Tools

• remember — governed write as the server's principal (classification, PII flag, source link, provenance signing, audit, encryption, indexing); pass supersedes to retire the memories it replaces • recall — policy-enforced recall; denied memories are never returned, or counted in the results or the receipt • explain_recall — ranking and freshness explanation for a recall; denied memories are never listed or counted in it • forget — crypto-shred the subject key and purge derived memories • evaluate_egress — check source spans before external-model egress • assess_faithfulness — check generated-memory claims against source spans • memory — Anthropic memory-tool-compatible /memories surface • health — readiness, warmed tenants, model names, key-custody mode

On the private runtime, versioned recall, context and facts responses, and the recall, context and facts agent tools, return no timing values. The older /recall address still returns the engine time for each request, and operators read the full stage timings from the metrics endpoint.

When the Jev reranker is attached, the explanation of a recall returns rerank stage timings.

Keep the MCP server local.

The MCP server in this guide runs over local stdio and should stay on the same machine as its client. Authenticated remote recall is a separate HTTP service: the Heartwood recall gateway supports per-organization bearer credentials, configurable network binding, and TLS.