Elastic Releases Atlas: Cognitive Science–Inspired Agent Memory System

Elastic has made Atlas publicly available—an agent memory architecture designed to tackle a fundamental problem in AI systems: how to efficiently surface relevant context from a user's entire interaction history without inflating token costs or degrading response quality.

The Problem with Long Context Windows#

Most AI agents today solve memory using context stuffing—packing historical conversation into the LLM's input. This approach falters quickly. While modern models support vast token windows, feeding entire interaction histories remains prohibitively expensive, increases latency, and triggers the "lost in the middle" phenomenon, where language models overlook information buried far from prompt boundaries. A million-token window, Elastic argues, is a temporary scratchpad, not a persistent memory system.

Three Memory Types from Cognitive Science#

Atlas applies principles from cognitive psychology, maintaining three distinct categories of agent memory:

• Episodic memory: records what happened—individual user interactions and observations
• Semantic memory: captures what's true—enduring facts and knowledge
• Procedural memory: stores what works—sequences of steps that solve recurring problems

Each type lives in its own Elasticsearch index with tailored lifecycle rules, allowing the system to evolve memories appropriately over time.

How Memories Form and Update#

New user inputs enter the system as episodic memories. Most fade naturally, but an LLM-driven consolidation process elevates key moments into lasting facts. During consolidation, the system extracts semantic memories (brief factual statements) and links them to supporting episodic evidence and any prior facts they replace. Concurrently, procedural memory grows—the LLM can author new "playbooks" (step-by-step problem solutions) and update success/failure metrics on existing ones, biasing future retrieval toward proven approaches.

Retrieval and Access#

Agents query across all memory indices through a hybrid search combining Reciprocal Rank Fusion merging of BM25 keyword search and semantic embeddings, with cross-encoder reranking. Document-level security ensures each user's memories remain isolated. Evaluation yielded a 0.89 Recall@10 score on question-answering tasks.