We built a persistent agent memory layer on Elasticsearch with 0.89 recall

A team built a persistent agent memory system on Elasticsearch achieving 0.89 recall at rank 10, using three indices (episodic, semantic, and procedural) with hybrid retrieval, RRF ranking, and per-user data-level security to prevent cross-tenant leaks. The system addresses limitations of context windows by providing long-term memory that persists across sessions, scales over years, and retrieves information by content, time, and user while handling contradictions through supersession rather than deletion. The full open-source implementation is available on GitHub.