Changelog

All notable changes to this project are documented here.
Format follows Keep a Changelog;
versioning follows Semantic Versioning.

[1.0.0] - 2026-08-10

First public release.

npx memgw start: local-first, zero-config startup. Creates ~/.memgw, generates
keys, binds to loopback. The server refuses to start without an auth key, and
requires a strong key when bound beyond loopback.
Three-layer store: raw events (90-day retention), distilled facts (forever),
Markdown topic notes in a git repo (forever). Configuration is centralised in
src/config.js with a documented precedence chain.
CLI: setup (one-command wizard: config, supervision, wires every agent found
on the machine), start, init, status, doctor, search, save,
forget, embed, watch, hooks.
Multi-agent capture. Native paths for Claude Code (hooks) and opencode (plugin),
transcript watcher for Codex CLI and any other JSONL-writing CLI, plus a generic
parser for unknown formats. MEMGW_CAPTURE_IGNORE excludes directories from
capture (chiefly the memgw checkout itself).
MCP server (Streamable HTTP, stateless) with five tools, reachable by header auth
or by path secret for web clients that cannot send headers.

SQLite FTS5 with porter stemming and diacritic-insensitive matching.
Optional semantic layer: hybrid BM25 + vector retrieval with reciprocal-rank
fusion. Vectors are Float32 BLOBs in the same SQLite file — no vector database,
no new dependencies. Toggle with memgw embed on|off|status; off by default,
and search falls back to BM25 whenever the embeddings API is unavailable.
Serve-time budget caps on /bootstrap (profile size, topic count).

Prompts in English and Vietnamese (MEMGW_PROMPT_LANG). Relative time
("yesterday") is resolved to absolute dates; assistant recaps of already-stored
facts are ignored; dedup defaults to "skip" for paraphrases and forbids lossy
merges (guarded by effectiveness test T4b — a real-LLM suite run before each release, not part of the keyless CI).
Reasoning models (gpt-5 family, o-series) supported: they receive
max_completion_tokens and no explicit temperature.

scripts/bench-locomo.mjs and scripts/bench-personamem.mjs: end-to-end recall
benchmarks through the real pipeline, with resume logs. Measured results in
docs/06-BENCHMARKS.md: LoCoMo 58.6% (BM25) / 66.4% (hybrid), PersonaMem 32k
59.6%.

Retention with three safety rails: only processed events are eligible, the newest
200 are always kept, and windows under 7 days are refused.
Backups: Litestream to S3-compatible storage, and an optional git push of the
notes directory.
deploy/com.memgw.plist (macOS launchd) and deploy/memgw.service (systemd) keep
the gateway alive; the bootstrap hook announces a dead gateway in the session
context instead of failing silently.
test/verify-docs.mjs fails the suite when documentation drifts from code.