Roadmap

Roadmap

Each phase is a Milestone; the tasks below are Issues. The mainline stays runnable — every later phase adds or swaps a pluggable implementation without tearing down the previous one.

Phase 0 — Scaffolding & decision lock (~0.5–1 day)

• Repo + FastMCP HTTP skeleton (/mcp) · create Neon · CREATE EXTENSION vector.
• Pin the embedding model (multilingual small), fix the dimension; confirm fastembed loads in-container.
• Auth / Store / Retriever interface stubs.

Definition of Done: MCP Inspector connects to /mcp with a bearer; the server can turn text into a vector.

Phase 1 — MVP: semantic memory (~2–3 days)

• Streamable HTTP + per-user bearer + Neon pgvector.
• remember (embed on write) / recall (vector NN) / list / forget + limit/cursor pagination.
• Chunk on write.

Definition of Done: Claude Desktop pointed at http://…/mcp with a token; recall("车") surfaces "丰田换机油"; data + vectors land in Neon.

Phase 2 — Hardening & shared validation (~1–2 days)

• Annotations, structured errors, pytest, namespaces (shared vs private).
• save_session / load_session, resources exposing memory://{namespace}.
• (optional) KeywordRetriever + hybrid (RRF).

Definition of Done: two different clients on the same Neon DB read each other's memories; core paths are tested.

Phase 3 — Launch (~1–2 days)

• Deploy (Cloud Run), enforce HTTPS, bearer keys via secret management.
• README + architecture diagram + demo GIF.
• Publish to PyPI → server.json → mcp-publisher publish; list on Glama / mcpservers.org.

Definition of Done: a stranger can configure a key + endpoint from the README and use it.

This week (minimal start)

1. Phase 0: repo + FastMCP HTTP skeleton + Neon (pgvector on) + embedding model chosen + three interface stubs.
2. Phase 1: remember (embed on write) + recall (vector) + per-user bearer working end-to-end, data + vectors in Neon.
3. Wire into Claude Desktop, store one fact with a token, recall it semantically.