Elastic open-sources Atlas agent memory system with cognitive science approach to long-term context

Elastic released Atlas, an open-source agent memory system built on Elasticsearch that addresses the long-term memory problem in agentic AI. Rather than attempting to fit years of interaction history into a fixed context window, Atlas maintains three types of memory grounded in cognitive science: episodic (what happened), semantic (what's true), and procedural (what works). The system integrates with agents via Model Context Protocol (MCP) and ensures per-user memory isolation.

Atlas stores episodic memories from user inputs with natural decay, consolidating some into durable semantic facts via LLM review. New semantic memories are stored with supporting evidence and previous facts they supersede. Procedural memory is maintained as playbooks (step-by-step problem-solving sequences) with success/failure counters to bias retrieval toward higher-performing solutions. The system retrieves memories using a hybrid approach: Reciprocal Rank Fusion over BM25 lexical search and Jina v5 semantic search, with results re-ranked via cross-encoder, all with document-level security ensuring queries only surface user-owned memories.

For agent builders, Atlas addresses a critical scaling problem: context window stuffing breaks down on cost, latency, and the "lost in the middle" effect where models ignore facts far from prompt edges. By separating long-term persistent memory from per-interaction context, agents can maintain coherent, multi-turn reasoning over years of interaction without bloating per-request token counts. The open-source release with MCP integration makes it reusable across agent frameworks.