memex

Personal cross-session, cross-harness (Pi + Codex) memory backend. Under construction: delivery phases 1–3 (DESIGN §11), validated on synthetic data only.

• docs/INSTALL.md — installing the server and the two plugins, settings, known limits
• docs/DESIGN.md — design (private, not in the public repository)
• docs/RESEARCH-landscape.md — survey of open-source memory projects and what was adopted
• dryrun/ — offline projection rules (spans, views, forks, canonical records), validated on local sessions
• prototype/ — sync protocol model
• server/ — the service: L0 sync (sync.ts), projection (project.ts), summaries and embeddings (derive.ts), search / zoom / grep (search.ts), forget (forget.ts), HTTP (http.ts), entry (main.ts)
• client/ — record encoding, the sync client and the mem_* tools shared by the plugins; cli.ts (import)
• plugins/pi/ — Pi extension; plugins/codex/ — Codex plugin (hooks, MCP server, build.ts)
• eval/ — question bank tooling

Tests: a Postgres 17 with pgvector and pg_trgm at MEMEX_PG_URL (default postgres://postgres:memex@127.0.0.1:55433/memex), then bun run test (server, client, plugins; the Pi end-to-end test needs pi on PATH) and bun test dryrun. The live Codex test is opt-in: see plugins/codex/e2e.test.ts.

Not implemented yet (stage 2 search contract, DESIGN §6.4)

• mem_search filters since / until / repo, and dates and repository on each result.
• A default of at most 8 results for the agent tool (the API takes limit, default 10, max 40).
• Task status on results and the "related sessions" grouping (they come from session cards and the summary forest, stage 3).
• mem_zoom above span level (session card → forest root → children), stage 3.
• /memex trace and /memex why <id> (the trace is written; the commands come with the plugins, stage 4).
• Trace retention (14 days [待标定]).
• The trace does not store the query as its own object (DESIGN §6.5): the query text is not kept at all, and the trace row is bound to the querying source and deleted with it on forget.
• The trace has no artefact versions (summary / embedding version per candidate) and no injection confirmation: returned and pool[].returned mean "assembled into the response to be delivered" — not delivered, not injected into a context; the plugins (stage 4) report what they inject. pool[].selected means chosen for the rerank set; rerank_skipped is why rerank scores were not used and does not say whether the request was sent (an endpoint error after sending also sets it).
• Regex mem_grep joins every readable record's text and relies on the 5 s statement timeout; its cost at real corpus size is not measured.