https://github.com/yantrikos/yantrikdb-server
Article
-
Purpose-built vector DB for AI agent memory with forgetting, consolidation, and contradiction detection
-
Built after ChromaDB recall quality degraded badly at ~5k memories
-
Uses half-life decay and LLM-assisted merging to prune stale/redundant facts
Discussion
-
Author confirms ChromaDB degradation was the direct motivation; recall became garbage at scale
-
Skepticism: contradiction detection without context is very hard — most facts aren’t cleanly atomic
-
Compared to mem0; one commenter stopped using mem0 because “facts” are too rigid for real tasks
-
Interest in consolidation loop details: random sampling + LLM to merge?
Discuss on HN