GoodMemory 0.8.2
Fixes
- Acquire SQLite writer admission before vector identity reads and bound WAL setup recovery to the existing one-second contention budget. Invalidate vector-table caches after rolled-back mutations.
- Coordinate same-process memory writes and projection invalidations without starving writeback.
- Validate observation source snapshots before maintenance generation, recall and provider calls; discard retrieval cues when their supporting facts change.
- Reject stale preference assertions using source evidence and chronology, and preserve personal attribution through embedded document containers and immutable source snapshots.
- Preserve adjacent numbered Chinese identifiers during retrieval and explicit English project assignments as literal facts.
Experimental Surface
- Add the opt-in, default-off
goodmemory/experimental/shadow subpath and an explicit JSON-configured example host. Advice remains read-only and is not connected to remember, conflict resolution or deletion. GoodMemory has no CognitiveHub runtime dependency.
Distribution
- Align npm, MCP, capability descriptors, setup guides and Kimi plugin pins to 0.8.2.
- Publish the exact locally prepared tarball, release manifest, deterministic evidence archive and plugin ZIP. The release workflow verifies those existing artifacts and npm integrity without rebuilding.
- Node.js >=20 and Bun >=1.3.14 remain required. This release makes no benchmark-quality claim and does not promise unbounded SQLite contention tolerance.
npm install -g goodmemory@0.8.2