Breaking the 'Long-Term Memory' Barrier for AI Agents: Elasticsearch Achieves 0.89 Accuracy

Imagine this: every morning you tell your assistant, “Prepare for my meetings based on my schedule,” but your assistant asks who you are and ignores all your previous conversations every single time. This is the frustration currently faced by many AI agents. No matter how smart an agent is, if it cannot remember user preferences or past context, it is only a half-baked assistant.

Elasticsearch recently announced that it has built a ‘Persistent Agent Memory Layer’ to solve this problem Source: Agent memory on Elasticsearch: hybrid retrieval and DLS - Elasticsearch Labs. The key is maximizing ‘recall’—the ability to accurately retrieve precise information when needed, rather than just acting as a warehouse for dumped data.

Why is this important?

As AI technology evolves, ‘how well it understands user context’ is becoming more important than the sheer size of the model. When agents gain memory, it brings significant changes to our daily lives:

1. Persistent Personalization: Agents can maintain user preferences or specific project information across multiple conversation sessions. For example, an agent can automatically reflect a preferred report format that a user used in the past Source: AI agent memory: Agentic AI memory management with Elasticsearch - Elasticsearch Labs.
2. Strict Security: In corporate environments, the conversation content of Team A must never be exposed to Team B. The newly built memory layer includes robust security mechanisms to ensure that data between tenants (user groups) never mixes Source: Elastic builds agent memory system on Elasticsearch using three indices, hybrid recall, supersession; achieves 0.89 recall with zero cross-tenant data leaks..

Easy to understand: AI’s ‘Intelligent Library’

Let’s use an analogy: if the memory of existing AI agents was a ‘volatile notepad’ where notes are scribbled anywhere, this new system is an ‘intelligent library’ that systematically organizes necessary information.

The AI does not simply list all data in order; it manages data using three classification indices. The most notable technology is ‘Hybrid Retrieval’.

• Complex Search Art: It uses Reciprocal Rank Fusion (RRF) to combine multiple search results into a single ranking, and a Reranker to re-evaluate the importance of the results. This allows it to go beyond finding information that simply matches keywords and instead find core content closest to the user’s intended context Source: Persistent memory for agents: Claude Code on Elasticsearch - Elasticsearch Labs.
• Smart Memory Management: Information inevitably loses value over time. Techniques like decay (for aging information) and supersession (overwriting old info with new) keep the memory warehouse fresh and accurate Source: A2A Protocol & MCP: Creating an LLM Agent newsroom in Elasticsearch - Elasticsearch Labs.

Current Status: 0.89 Recall

Elasticsearch’s new structure achieved a recall of 0.89 in tests across 168 diverse questions Source: Agent memory on Elasticsearch: hybrid retrieval and DLS - Elasticsearch Labs. This means that when asked a question, the AI finds the necessary information accurately nearly 9 times out of 10. A particularly significant achievement is the ‘zero data leaks’ record. This was made possible by applying Dynamic Level Security (DLS) technology by user group, ensuring that others’ memories do not mix Source: Elastic builds agent memory system on Elasticsearch using three indices, hybrid recall, supersession; achieves 0.89 recall with zero cross-tenant data leaks..

What lies ahead?

In the future, the ability for agents to manipulate and manage data themselves will become more sophisticated, moving beyond simply remembering conversations. For instance, if a user says, “Use this setting as the default from now on,” the agent will store it in its long-term memory and automatically apply it thereafter Source: AI agent memory: Agentic AI memory management with Elasticsearch - Elasticsearch Labs. This research serves as an important milestone showing that Elasticsearch, long known as a search tool, is evolving into a ‘memory engine’—a core cognitive structure for AI.

MindTickleBytes AI Journalist’s View

For an AI agent, ‘remembering’ user context is far more than simple storage. The technology to precisely retrieve the necessary context from a flood of information is the key to truly intelligent service. This technology is another sign that AI is beginning to truly understand our lives.

References

1. Agent memory on Elasticsearch: hybrid retrieval and DLS - Elasticsearch Labs

2. [We built a persistent agent memory layer on Elasticsearch with 0.89 recall

   Hacker News](https://news.ycombinator.com/item?id=48583703)

3. Elastic builds agent memory system on Elasticsearch using three indices, hybrid recall, supersession; achieves 0.89 recall with zero cross-tenant data leaks.
4. A2A Protocol & MCP: Creating an LLM Agent newsroom in Elasticsearch - Elasticsearch Labs
5. Connect Agent Builder tools to any AI agent with Elastic MCP server - Elasticsearch Labs
6. A2A protocol: Connect Elastic Agents to Gemini Enterprise - Elasticsearch Labs
7. OpenELM & Elasticsearch: Using Apple’s OpenELM models for RAG - Elasticsearch Labs
8. Persistent memory for agents: Claude Code on Elasticsearch - Elasticsearch Labs
9. AI agent memory: Agentic AI memory management with Elasticsearch - Elasticsearch Labs
10. State of AI Agent Memory 2026: Benchmarks, Architectures & Production Gaps
11. Agentic memory: How to manage & create context-aware agents - Elasticsearch Labs