Elastic’s Search Labs describes Atlas, a demo architecture for persistent AI agent memory that combines Elasticsearch hybrid retrieval, reranking, tenant isolation controls, and MCP support. The accompanying GitHub materials frame the project as a reference implementation rather than a production product.
Elastic’s Search Labs has published a technical walkthrough of an Elasticsearch-based persistent memory layer for AI agents, describing a multi-tenant design that reports 0.89 recall and no cross-tenant leakage in its evaluation.
The Elastic Search Labs post, titled “How we built a persistent agent memory layer on Elasticsearch with 0.89 recall and zero tenant leaks,” presents an architecture for giving AI agents access to stored memories across sessions. The article describes the system as a way to persist and retrieve information for agent clients using Elasticsearch, hybrid retrieval, reranking, document-level security, and a Model Context Protocol endpoint.
The related GitHub repository, noamschwartz/atlas-memory-demo, identifies the project as Atlas, an Elasticsearch-based agent memory demo. According to the repository README, Atlas separates memory into episodic, semantic, and procedural categories, giving the demo a structure for storing different kinds of agent-relevant information.
The GitHub README says Atlas uses a combination of BM25 and dense retrieval with Jina v5 embeddings, joined through Reciprocal Rank Fusion. The project’s ATLAS.md guide also says the demo uses Elasticsearch semantic_text auto-embedding and RRF-based hybrid recall.
Elastic’s post reports a recall result of 0.89. The ATLAS.md source gives additional evaluation context, stating that the project’s pass criteria require Recall@10 of at least 0.85 and cross-tenant leakage of 0. The same guide describes conflict resolution and Claude-based consolidation as part of the demo’s approach to maintaining memory quality over time.
Those details matter because persistent memory systems for agents need to retrieve useful prior context without mixing data between users. Elastic’s material positions hybrid retrieval and reranking as the relevance layer, while the GitHub documentation presents isolation checks as a required evaluation criterion.
The Atlas README says the demo supports per-user isolation through user_id filtering or optional document-level security keys. Elastic’s Search Labs post also highlights document-level security isolation as part of the multi-tenant architecture.
The sources do not establish that Atlas is a finished commercial service. Instead, the public repository and project guide present it as a demo and reference implementation showing how Elasticsearch features can be assembled for agent memory use cases.
Elastic’s article says the memory layer exposes an MCP endpoint for agent clients. The Atlas README also lists MCP support, indicating that the demo is designed to connect with tools and agents using the Model Context Protocol.
That design choice reflects a broader trend in AI infrastructure: memory services are increasingly being treated as separate components that agents can call, rather than as logic embedded inside a single application. Based on the supplied sources, Atlas demonstrates one such approach using Elasticsearch retrieval, multi-tenant controls, and memory-type separation.
For teams evaluating persistent memory for agents, the useful takeaway is not that Atlas guarantees production outcomes, but that Elastic and the Atlas documentation provide a concrete, testable pattern: separate memory indices, hybrid retrieval, evaluation thresholds for recall, and explicit tenant-isolation checks.
The article describes the system as a way to persist and retrieve information for agent clients using Elasticsearch, hybrid retrieval, reranking, document level security, and a Model Context Protocol endpoint.
The related GitHub repository, noamschwartz/atlas memory demo , identifies the project as Atlas, an Elasticsearch based agent memory demo.
According to the repository README, Atlas separates memory into episodic, semantic, and procedural categories, giving the demo a structure for storing different kinds of agent relevant information.
Continue reading
