Elastic built a persistent agent memory layer on Elasticsearch with 0.89 recall
We built a persistent agent memory layer on Elasticsearch with 0.89 recall
Elastic open-sourced an agent memory architecture using three Elasticsearch indices for episodic, semantic, and procedural memory. Hybrid retrieval with a reranker hits R@10 of 0.89 across 168 questions, with zero cross-tenant leaks via per-user DLS. The post doesn't disclose latency or cost figures, but it details design choices: supersession over deletion for audit trails, decay so older facts don't outrank fresh ones, and consolidation of raw events into durable facts. Full implementation is on GitHub.
Why it matters: Elastic open-sourced an agent memory design that splits events, user facts and procedures into three indexes, hitting 0.89 recall at 10 on 168 test questions.