TMCRA 1.0.0-rc.1 — local-first memory workspace

Unified major release candidate

• Nine MCP tools: recall, ingest, turn prepare/commit, reconciliation, get/wait for job, session controls, and chat-confirmed source feedback.
• Automatically discover the private Windows local-memory identity after installing the standalone local runtime. Local mode requires no TMCRA cloud account or TMCRA server.
• Invalid/incomplete local setup fails closed; stale cloud connections are blocked. Hosted/self-hosted endpoint configuration remains an explicit option.

Install

Open the attached MCPB in a compatible host, or install the attached Python wheel/sdist (1.0.0rc1). The MCP adapter uses a host-provided Python/uv runtime; it does not embed a Python interpreter or model weights.

For local memory, use the Windows x64 one-click runtime, then restart the MCP host. Initial environment and model downloads require internet access. Memory execution and storage then use the user's computer.

Validation and limitations

47 tests passed locally. Python 3.10/3.12 on Windows and Linux passed in CI.

The full-local runtime is a preview: synthetic CPU ingest and raw recall passed; complex evidence compilation timed out. Organizer/full-service restart and higher-tier hardware acceptance remain pending. A cloud-hosted Agent model may receive retrieved memory; making the entire Agent offline also requires a local host model.

The MCPB SHA-256 matches the committed registry metadata: 714d649a527d55498ce9bcb4f0981934848c5d5bde2398f4cec0243128fb5860.