Amazon OpenSearch Memory Gives AI Agents Persistent Recall

Glowing translucent robot sending streams of light into a floating wall of data cards
AI-generated illustration

Ask an AI agent the same question on Monday and again on Friday, and it often answers as if the two of you had never met. On November 18, 2026, AWS announced that Amazon OpenSearch Memory is generally available, a serverless capability meant to fix exactly that by giving agents persistent, searchable memory, according to the AWS announcement.

What OpenSearch Memory does

OpenSearch Memory is a new part of Amazon OpenSearch Service. It stores an agent’s interactions as searchable events, derives useful memories from them, and retrieves relevant context across sessions. In AWS’s words, agents can recall prior decisions, earlier interactions and user preferences without replaying the full conversation history to the model.

That last detail matters in practice. Feeding a long transcript back into a model on every turn is slow and expensive, and it eventually runs into context limits. A memory layer that pulls back only what is relevant is a different design: the agent asks for what it needs, and the store answers.

How the memory lifecycle works

AWS says the service keeps the original events verbatim and makes them searchable within seconds. In the background it asynchronously derives structured memories from those events. It then manages the full lifecycle: extraction, embedding, indexing and consolidation, which means adding, updating or skipping a memory as new information arrives. Retrieval can use keyword, semantic or hybrid search.

Storing the raw events alongside the derived memories gives builders two views of the same history. The verbatim record is there for audit and replay, while the consolidated memories are the compact, query-friendly layer an agent actually reads.

Serverless, with controls

Customers control access through AWS Identity and Access Management (IAM) for each memory store, with actor and session scoping. They can also set expiration periods for both events and memories, so data does not have to live forever. Because the capability is serverless, AWS notes there are no clusters to provision, no instance types to choose and no indexes to create. Capacity for extraction and search scales automatically.

For teams that have been stitching together separate extraction, embedding, indexing, consolidation and retrieval pieces for agent memory, the pitch is to collapse that into one managed service.

Availability

Amazon OpenSearch Memory is available today in US East (N. Virginia and Ohio), US West (Oregon), Europe (Frankfurt) and Asia Pacific (Tokyo). Pricing is published on the OpenSearch Service pricing page, and AWS points to technical documentation and a launch blog for getting started.

What it means

Agent memory is moving from a do-it-yourself pattern to a managed building block. That lowers the barrier for teams that want agents to feel continuous, such as support assistants that remember a customer’s history or internal copilots that recall earlier decisions. It also raises questions every team should answer for itself: what should an agent be allowed to remember, for how long, and who can read it. The per-store IAM scoping and expiration settings give a starting point, but the policy choices still belong to the organization.

It is also worth keeping expectations measured. The announcement describes capabilities, not benchmark results, so how well derived memories hold up on messy real-world conversations is something builders will need to test.

Bottom line

OpenSearch Memory turns long-term recall for AI agents into a serverless service with search, lifecycle management and access controls built in. If agents are going to be useful across days rather than minutes, a layer like this is likely to become standard plumbing.

Sources