Elastic Open-Sources Atlas Agent Memory Based on Cognitive Science

2026-06-30 · Source: InfoQ · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, quick

Summary

Elastic has open-sourced Atlas, an Elasticsearch-based system designed to provide long-term memory for AI agents interacting with users over extended periods. Released on June 30, 2026, Atlas addresses the scalability and "lost in the middle" issues associated with stuffing entire interaction histories into LLM context windows. The system implements three cognitive science-inspired memory types—episodic, semantic, and procedural—each managed in separate Elasticsearch indices. Episodic memories capture events, semantic memories store consolidated facts identified by an LLM, and procedural memories track "playbooks" with success/failure metrics. Agents access these memories via a hybrid query combining BM25 lexical search and Jina v5 semantic search, re-ranked by a cross-encoder, achieving 0.89 Recall@10 in question-answering. Document-level security ensures per-user memory isolation.

Key takeaway

For AI Engineers designing conversational agents that require persistent, scalable memory across long user interactions, Elastic's open-sourced Atlas offers a robust architectural blueprint. You should evaluate its cognitive science-inspired three-memory-type approach and hybrid search strategy to overcome context window limitations and "lost in the middle" effects. Consider adopting its Elasticsearch-based design for managing diverse memory types and ensuring per-user data isolation, potentially saving significant development time.

Key insights

Atlas provides scalable, long-term agent memory by categorizing information into episodic, semantic, and procedural types based on cognitive science.

Principles

• Agent long-term memory benefits from distinct episodic, semantic, and procedural stores.
• LLMs can consolidate transient episodic events into durable semantic facts and procedural playbooks.
• Hybrid search combining lexical and semantic methods improves memory retrieval accuracy.

Method

Atlas stores user inputs as episodic events, uses an LLM to consolidate these into semantic facts and procedural playbooks, and retrieves memories via a hybrid RRF query over BM25 and Jina v5, re-ranked by a cross-encoder.

In practice

• Implement per-user memory isolation using document-level security.
• Employ Reciprocal Rank Fusion for combining diverse search results.
• Consider Elasticsearch for scalable, multi-modal memory storage with advanced features.

Topics

• Atlas
• Agent Memory
• Elasticsearch
• Large Language Models
• Reciprocal Rank Fusion
• Cognitive Science

Code references

• noamschwartz/atlas-memory-demo

Best for: AI Engineer, Machine Learning Engineer, AI Architect

Related on AIssential

• Memory Systems for Long-Running Agents: Episodic to Procedural — Long-running agents require structured, retrieval-based memory systems to overcome context window limitations and scale economically.
• APEX-MEM: Agentic Semi-Structured Memory with Temporal Reasoning for Long-Term Conversational AI — APEX-MEM enhances long-term conversational AI through structured memory, temporal grounding, and intelligent retrieval.
• Agent Memory — Agent memory is complex, requiring diverse memory types and intelligent management beyond simple context window extension.
• Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents — Memory retrieval for LLM agents is more effective as an active, reconstructive process over a graph.
• Designing Memory for AI Agents: Inside Linkedin’s Cognitive Memory Agent — LinkedIn's CMA provides a stateful memory layer for AI agents, enabling continuity, personalization, and adaptation across interactions.
• 30 FREE Tutorials to Build AI Agents With Real Memory Fast! — The repository provides a practical, comprehensive guide to agent memory techniques, from basic to production-ready systems.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by InfoQ.