We built a persistent agent memory layer on Elasticsearch with 0.89 recall
7 hours ago
- #Elasticsearch
- #Agent Memory
- #Hybrid Retrieval
- Agent Builder is now GA, with a memory system built on Elasticsearch featuring three indices for episodic, semantic, and procedural memory.
- The recall pipeline uses hybrid retrieval with BM25 + Jina v5 dense vectors and RRF, followed by a cross-encoder reranker for better accuracy.
- Memory writing includes refresh=True for same-turn visibility, and consolidation promotes episodic events into durable semantic facts and procedural playbooks.
- Supersession handles contradictions by hiding old facts but keeping them for audit, and multi-tenant isolation is enforced via Elasticsearch DLS.
- Time decay and use-count scoring adjust retrieval rankings, and the system integrates shared catalog data with a soft priority for user memory.
- The architecture supports any agent via MCP, with tools for recall, write, and forget, and eval metrics show R@10 averages 0.89 with zero cross-tenant leaks.