Last updated: November 19, 2026
Amazon OpenSearch Memory is a serverless memory service for AI agents that AWS made generally available on November 18, 2026. Your app writes conversation turns as events, the service extracts facts and summaries with Claude Haiku 5.5 or Opus 5.5, and your agent searches those memories with keyword, semantic or hybrid queries.
Contents
- What Amazon OpenSearch Memory does
- Spaces, events, memories and strategies
- How to set up your first space
- Extraction triggers and consolidation
- Pricing and Regions
- Quotas and limits
- Memory search vs. full history
- Limitations and tips
- Frequently asked questions
What Amazon OpenSearch Memory does
AWS announced the launch in a What’s New post dated November 18, 2026. It describes OpenSearch Memory as a new serverless capability in Amazon OpenSearch Service that gives agents long-term memory without clusters, instance types or indexes to manage.
The developer guide explains the core idea. The service handles storage, indexing, extraction and the model calls extraction needs. You do not request model access, run a vector database or write extraction prompts for common cases.
Spaces, events, memories and strategies
- Space: the AWS resource that stores events and memories and defines which memory strategies run.
- Event: an original interaction your application writes, stored verbatim. Events are readable as soon as the write returns and searchable within seconds.
- Memory: a fact, preference or summary derived from events in the background. Each memory lists its
sourceEventIds, so you can trace it back to the conversation. - Actor and session: every event and memory belongs to an actor (the user) and a session (one conversation). Searches are scoped to one actor, and the
osm:actorIdIAM condition key can enforce that.
The memory strategies page lists three strategy types. A space can run up to 10 strategies, each extracting from the same events.
| Strategy type | What it extracts | Example from AWS docs |
|---|---|---|
SEMANTIC |
Facts about the actor worth keeping across sessions | “Prefers window seats.” |
SUMMARY |
A summary of each session, updated as the session continues | A short account of a laptop comparison and the final choice |
CUSTOM |
Whatever your extraction prompt describes (up to 30,720 characters) | “Order 4471: delayed, refund issued 2026-11-02.” |
When your application already knows a fact, such as a preference from a CRM, it can call CreateMemory to write the memory directly instead of waiting for extraction.
How to set up your first space
The getting started tutorial uses the AWS CLI, where commands live under aws opensearchmemory. The SDK service ID is OpenSearchMemory, and the launch blog’s Python walkthrough needs Boto3 1.44.4 or later.
- Grant your IAM identity the actions you need, such as
osm:CreateSpace,osm:GetSpace,osm:CreateEvent,osm:SearchMemoriesandosm:DeleteSpace. - Create a space with
aws opensearchmemory create-space, for example with oneSEMANTICstrategy and oneSUMMARYstrategy. In the console, choose Memory, then Spaces, then Create Space; Easy create adds the same two strategies. - Wait for the space to become
ACTIVEwithaws opensearchmemory wait space-active. AWS says this usually takes a few minutes. - After each agent turn, send the conversation with
create-event, including anactorIdand a stablesessionId. - Before each response, call
search-memoriesfor that actor and add the results to the model prompt. Ahybridquery can combine amatchquery onmemory_textwith aneuralquery. - When you finish testing, delete the space with
delete-space. Deletion is permanent.
Extraction triggers and consolidation
A strategy does not extract after every message. It runs when the first of three triggers is reached: extractEveryNMessages (default 12), extractEveryNTokens (no default) or idleTimeoutSeconds (default 60 seconds). Raise the idle timeout if users pause longer than 60 seconds.
Consolidation is on by default. When a new memory refines or contradicts a stored one, the service replaces it; duplicates are skipped. The docs give the example of “Lives in Seattle” being replaced by “Moved to Austin.” Because an update creates a new memoryId, do not store memory IDs long term.
Pricing and Regions
The OpenSearch Service pricing page says you pay for memory records stored each month, retrieval requests (search, get and list, at one rate for lexical, semantic and hybrid search), and model tokens for managed extraction. There is no ingestion charge and no minimum fee. Each record is billed for at least one month.
Managed extraction uses Claude Opus 5.5 or Claude Haiku 5.5. AWS says token rates match Amazon Bedrock public pricing for the same model and geography with no markup. Memories you write directly do not incur extraction charges. AWS’s Price List API file for US East (N. Virginia) lists retrieval at $0.04 per 1,000 requests. Check the pricing page for record storage and token rates in your Region. For the model, see our Claude Haiku 5.5 guide.
The developer guide and the endpoints and quotas page list four Regions: US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo) and Europe (Frankfurt). The What’s New post also names US East (Ohio), which does not appear in the endpoint table on November 19, 2026.
Quotas and limits
| Limit | Default | Adjustable |
|---|---|---|
| Spaces per account per Region | 50 | Yes |
| Memory strategies per space | 10 | No |
| Event retention | 1 to 365 days | No |
| Memory retention | Up to forever | No |
| Messages per conversational event | 100 | No |
| Text per event | 5 MiB | No |
| Extraction tokens per minute, per account per Region | 400,000 | Yes |
| Memory content per CreateMemory request | 16,000 characters | No |
| CreateEvent / SearchMemories rate | 50 / 30 requests per second | Yes |
Memory search vs. full history
In the launch blog, AWS reports a LongMemEval-S test. Claude Opus 5.5 answered from three kinds of context. These are AWS’s own single-run results.
| Context given to the model | Answer accuracy | Context tokens per question |
|---|---|---|
| OpenSearch Memory, top 30 memories and events (hybrid search) | 93.0% | About 13,500 |
| BM25, top 30 raw conversation turns | 88.6% | About 22,700 |
| Full conversation history | 95.2% | About 167,000 |
The trade-off is a small accuracy drop for far fewer input tokens. For other approaches, see our guides to Codex memory and Mistral managed agents. Amazon’s agent tooling also includes the AgentCore Web Search API and AgentCore Web Answer.
Limitations and tips
- Memories lag events. A new user’s memory search can return nothing at first; read recent events with
ListEventsto resume a session. - Encryption is set at creation. Data is encrypted with an AWS owned key by default; a customer managed AWS KMS key can only be set when you create the space.
- No FIPS endpoints. OpenSearch Memory does not offer them.
- Review agent-written memories. AWS suggests tagging them with metadata such as
writtenBy.
Frequently asked questions
What is Amazon OpenSearch Memory?
It is a fully managed, serverless memory service for AI agents in Amazon OpenSearch Service. Your application sends conversations as events, the service extracts facts and summaries as memories, and your agent searches those memories to recall what it learned across sessions.
When did Amazon OpenSearch Memory become generally available?
AWS announced general availability on November 18, 2026.
Which AWS Regions support OpenSearch Memory?
The developer guide and endpoint table list US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo) and Europe (Frankfurt). The launch announcement also names US East (Ohio), so confirm Ohio in the console or endpoint table before you plan around it.
How is OpenSearch Memory priced?
AWS charges for memory records stored each month, for retrieval requests, and for model tokens used by managed extraction. Ingestion is not charged and there is no minimum fee. Records are billed for at least one month, and memories you write directly skip extraction charges.
Which models extract memories?
Managed extraction uses Anthropic Claude Haiku 5.5 or Claude Opus 5.5, which you select. The launch blog says Haiku 5.5 is the default. AWS says extraction token rates match Amazon Bedrock public pricing for the same model and geography, with no markup.
Sources
- AWS What’s New: Amazon OpenSearch Memory is now generally available (November 18, 2026)
- AWS Big Data Blog: Introducing Amazon OpenSearch Memory
- Amazon OpenSearch Service Developer Guide: What is Amazon OpenSearch Memory?
- Amazon OpenSearch Service Developer Guide: Endpoints and quotas for OpenSearch Memory
- Amazon OpenSearch Service pricing
Image: AWS logo, public domain, via Wikimedia Commons. AWS and Amazon OpenSearch Service are trademarks of Amazon.com, Inc. or its affiliates.