Memory-Facade

A thin, stateless MCP server that gives every coding agent (Claude Code, Codex, OpenCode, Hermes) one curated entry point to shared memory — automatic bank/tag routing, URL-ingest → linked article, session → documentation set, and consistency tooling (dedupe / reroute).

It orchestrates Hindsight (facts) and LightRAG (corpora). It does not reimplement storage and never creates banks.

See ~/Projects/common-memory/docs/memory-facade-architecture.md for the design and the 2026-08-18 content baseline.

Run (stdio MCP server)

uv run python -m mf.server

Deployed as an MCP server in the msm-ai-gateway LiteLLM MCP runtime, routed as /memory-facade/mcp (same pattern as lightrag in config/litellm.yaml).

Test

env -u PYTHONPATH uv run --extra dev pytest -q

Note: on Sergey's host the shell exports a PYTHONPATH pointing at the Hermes venv. Unset it (env -u PYTHONPATH …) before running here, otherwise the wrong pydantic/pydantic_core gets imported and collection fails.