· via dev.to (home feed)
Amazon OpenSearch Memory hits general availability as managed memory for AI agents
Amazon's managed memory service for AI agents is now generally available, storing what an agent learns and returning it as searchable memories at a fraction of the token cost of full conversation history.
What launched
Amazon has made OpenSearch Memory generally available, a managed service that keeps track of what an AI agent learns and serves it back as searchable memory in later sessions. According to a dev.to write-up of the launch, AWS announced general availability on 18 November 2026, with the service built on Amazon OpenSearch Serverless and offered in US East (N. Virginia), US West (Oregon), Europe (Frankfurt) and Asia Pacific (Tokyo).
The problem it addresses is straightforward: stuffing an entire conversation into a prompt only works while the history fits inside the context window and the budget. OpenSearch Memory treats memory as an API-backed service instead. An application writes each turn, the service works out which details are worth persisting, and the next session pulls back a short list of relevant context.
Spaces, events and extraction
The central resource is a space, which behaves much like a database for an application: a single store with its own endpoint, strategy configuration and retention rules. Application code calls CreateEvent after each turn with the message payload. Events are kept verbatim, indexed for keyword, semantic and hybrid search, and queryable straight away.
Deriving memories from those events happens asynchronously in the background. Strategies attached to the space — facts, preferences and session summaries ship in the default set, with up to 10 per space — all read from the same event stream. A consolidation pass compares each candidate memory with what is already stored, then adds, updates or discards it, which keeps the store from filling up with contradictions.
Extraction runs through a Bedrock model that defaults to Claude Haiku 5.5 and can be swapped per space via extractionModelId. The dev.to report notes that extraction and embedding costs are folded into the service price, and that switching models later re-derives memories from the original event log.
Actors, retention and encryption
Each request identifies an actor, the end user the data belongs to, and an optional session ID groups the events of one conversation, with reads and searches restrictable to a single session. An osm:actorId condition key on the agent's IAM role constrains which actors that role can act for, so separation between customers rests on IAM policy.
Retention is configured per space: events persist for 1 to 365 days, and memories anywhere from 1 day to indefinitely, with the launch material showing a worked example of 90 days for events and 365 for memories. Data is encrypted through AWS KMS, a customer-managed key can be supplied but only at space creation, and API activity is logged to CloudTrail.
The benchmark numbers
AWS tested the full workflow on LongMemEval-S, a public benchmark of 500 questions about information spread across roughly 50 chat sessions for a single user, from work by Wu and co-authors presented at ICLR 2025. Conversations were written in as events, default fact and summary strategies produced memories, and each question retrieved the top 30 memories and events through hybrid search. Claude Opus 5.5 produced and graded the answers, with Amazon Titan Text Embeddings generating the vectors.
Three ways of filling the prompt were compared. Managed memory answered 93.0% of questions using roughly 13,500 input tokens per question; BM25 keyword search over raw turns scored 88.6% on about 22,700 tokens; and the complete conversation history reached 95.2% on roughly 167,000 tokens.
The dev.to write-up highlights the caveat AWS attaches to its own table: one run per configuration and a 95% confidence range of two to three points either way, which means the 2.2-point gap between managed memory and full history falls inside the reported band. The firmer signals are the measured token counts — managed memory used about one twelfth of the input tokens of full history — and the 4.4-point advantage over keyword search over raw turns.
Why it matters
Agent memory is shifting from something every team hand-rolls into rented infrastructure, and AWS is pricing the compromise explicitly: near-parity accuracy within the reported confidence band, at roughly one twelfth of the input cost, with context headroom freed up for the actual task. The constraints worth weighing before adoption are operational rather than algorithmic. Four regions leave no option for stricter data-residency requirements, such as keeping all data inside France. Memory quality inherits whichever Bedrock model extracts facts from your transcripts. And the retention and actor-mapping decisions made before the first write determine how painful it becomes when a customer later asks for their data to be deleted.
- #aws
- #opensearch
- #ai-agents
- #agent-memory
- #cloud