At ~10K documents, vector similarity becomes noise — everything matches everything. And no existing system tracks identity, relationships, emotional context, or learned procedures.

OpenMemory can either replace your memory system or layer on top of it to allow you to toggle it off and on and actually compare memory results in real time.

OpenMemory is 7 layers on one Postgres database: 1. Sensory Buffer (6-stage input pipeline) 2. Semantic (Apache AGE graph + pgvector) 3. Episodic (experiences with emotional arcs) 4. Identity (agent values, beliefs, purpose) 5. Relational (person models with trust vectors) 6. Procedural (learned workflows) 7. Meta-Memory (consolidation engine)

The real differentiator is sleep cycles — session, nightly, and weekly passes that consolidate memories, extract lessons, resolve contradictions, and decay stale data. Like how your brain processes during sleep.

Tech: TypeScript, PostgreSQL 18, Docker-ready. The demo in docs/DEMO.md shows the before/after of a sleep cycle on the same data.

I'd love technical feedback — especially on the consolidation engine design. PRs welcome.

Can you add the URL?

[dead]