How we built Engrava: from cognitive-architecture research to a production library

Short summary

Engrava is a typed knowledge graph for agent memory stored in a single SQLite file, using deterministic consolidation instead of LLM rewrites. Its 'dreaming' cycle scores thoughts on recency, staleness, confirmation, confidence, and frequency before promoting them, drawing on cognitive-architecture research. The authors honestly report that consolidation doesn't move retrieval benchmarks, arguing the benchmarks measure the wrong thing.

• •Engrava stores agent memory as a typed graph in SQLite with hybrid search (vector + keyword + recency + priority)
• •Deterministic 'dreaming' consolidation cycle inspired by sleep-memory research promotes only re-confirmed facts
• •Authors acknowledge consolidation doesn't improve benchmark scores, suggesting benchmarks miss what matters

Generated with AI, which can make mistakes.