Elastic’s Atlas is an MCP-compliant persistent memory system that replaces temporary context windows with a human-like cognitive architecture to provide AI agents with scalable, multi-year memory.
Elastic has officially open-sourced Atlas, a persistent memory system built on top of Elasticsearch designed to provide long-term, scalable memory for AI agents. Atlas connects with AI agents using the Model Context Protocol (MCP).
Standard large language models (LLMs) rely on massive context windows (e.g., 1 million tokens) to remember user interactions. Elastic notes that context windows function like temporary “scratchpads” rather than a real memory system. Stuffing multi-year chat histories directly into a prompt breaks down due to high computational costs, processing latency, and the “lost in the middle” effect (where LLMs ignore facts placed far from the prompt’s edges).
Atlas acts as a long-term, persistent store that survives the end of a chat session, scales to years of data, and allows agents to fetch facts filtered by content, time, and specific user ID. Elastic Atlas architectures AI agent memory by mirroring human cognitive science, routing data into three distinct Elasticsearch indices that manage unique data lifecycles:
-
Episodic Memory (“What Happened”): Captures individual chronological logs of every user input and transaction event.
-
Semantic Memory (“What’s True”): Stores consolidated, durable facts. An LLM periodically runs in the background to review decaying episodic memories, extract generalised truths into short sentences, and map them to the previous facts they supersede.
-
Procedural Memory (“What Works”): Tracks action workflows. Background consolidation generates step-by-step problem-solving instructions and updates success/failure counters to bias the agent toward more effective workflows over time.
Agents access Atlas via a single unified query that executes a hybrid search. It combines BM25 lexical keyword search with Jina v5 semantic vector search using Reciprocal Rank Fusion (RRF), and refines the final payload using a cross-encoder reranker.
When evaluated on its question-answering capabilities, Atlas achieved a high accuracy benchmark score of 0.89 Recall@10. To ensure multi-tenant security, Atlas relies on Elasticsearch’s native Document-Level Security (DLS) to strictly isolate memories, ensuring queries only fetch data matching that specific user’s credentials.
