Amazon Bedrock AgentCore Memory
Amazon Web ServicesAmazon Bedrock AgentCore Memory is a managed AWS service that keeps an agent's conversation events as short-term memory and runs memory strategies over them to extract long-term records: facts, user preferences, session summaries and episodes. New information is consolidated into existing records, and the agent recalls them by semantic search within an actor's or session's namespace. AWS offers it as one of the Amazon Bedrock AgentCore services.
Openness
1 high confidence- license
- Proprietary (AWS managed service used under the AWS Service Terms)
- source
- closed (extraction, consolidation and retrieval run inside AWS, and no repository publishes the service)
- clients
- Apache-2.0 (the bedrock-agentcore Python SDK, a client for every AgentCore service, calls the hosted APIs)
- deployment
- cloud-only (billed per event ingested, stored and retrieved and per long-term record stored and retrieved)
AgentCore Memory runs only inside AWS, which bills it by the event and the memory record, and the code that extracts and consolidates memories is not published or offered for self-hosting. The Python SDK AWS publishes under Apache-2.0 calls the hosted service without letting anyone run it.
- https://aws.amazon.com/bedrock/agentcore/pricing/ recorded 2026-11-20
Under Memory: short-term memory is priced on "ingestion of raw events, storage of those events over their time to live (TTL), and retrieval of events"; long-term records on "the number of memories processed and stored each month and the number of memory record retrieval calls".
- https://aws.amazon.com/service-terms/ recorded 2026-11-20
The AWS Service Terms state that "The Service Terms below govern your use of the Services", so AgentCore is used under AWS service terms rather than a published software license.
- https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory.html recorded 2026-11-20
The developer guide says "AgentCore Memory is a fully managed service that gives your AI agents the ability to remember past interactions" without "the need to build or manage complex infrastructure".
- https://pypi.org/pypi/bedrock-agentcore/json recorded 2026-11-20
PyPI metadata for bedrock-agentcore: summary "An SDK for using Bedrock AgentCore", classifier "License :: OSI Approved :: Apache Software License", homepage github.com/aws/bedrock-agentcore-sdk-python.
Adoption
not assessedAWS publishes no customer, memory or request count for AgentCore Memory itself; its documentation, pricing page and feature announcements were searched for one. The bedrock-agentcore SDK's downloads cannot stand for it, because that one package serves every AgentCore service, not memory alone.
- https://aws.amazon.com/about-aws/whats-new/2026/09/agentcore-memory-direct-ingest/ recorded 2026-11-20
The announcement of the IngestData API for direct long-term memory ingestion describes the feature and gives no customer, record or request count.
- https://pypi.org/pypi/bedrock-agentcore/json recorded 2026-11-20
The bedrock-agentcore package describes itself as "An SDK for using Bedrock AgentCore" and its README opens "Deploy your local AI agent to Bedrock AgentCore with zero infrastructure", covering all of AgentCore.
- https://pypistats.org/api/packages/bedrock-agentcore/recent recorded 2026-11-20
7,382,429 downloads of bedrock-agentcore in the trailing 30 days, across every AgentCore service.
Capability
4 medium confidenceAgentCore Memory extracts records from raw events with a managed model, consolidates them against what is already stored, and adds reflections across past episodes, keeping facts, preferences, summaries and episodes as separate kinds of record. Records leave only when a retention period set on the memory runs out; the service does not itself judge a memory stale or weaken it with age.
- https://docs.aws.amazon.com/bedrock-agentcore-control/latest/APIReference/API_CreateMemory.html recorded 2026-11-20
CreateMemory requires eventExpiryDuration, "The number of days after which memory events expire, between 7 and 3650 days"; each strategy in the response carries recordExpiryDurationDays.
- https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/built-in-strategies.html recorded 2026-11-20
Built-in strategies run "Extraction", "Consolidation – Determines whether to write useful information to a new record or an existing record" and "Reflection – Insights are generated across episodes"; the strategies are semantic, user preference, summary and episodic.
- https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/episodic-memory-strategy.html recorded 2026-11-20
The episodic strategy "automatically detects episode completion" and stores episodes and reflections under namespaces such as /strategy/{memoryStrategyId}/actor/{actorId}/session/{sessionId}/.
- https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-terminology.html recorded 2026-11-20
Defines a namespace as "a structured path used to logically group and organize long-term memories" and says "A unique actorId maintains that memory records are correctly associated with the individual or system."
- https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-types.html recorded 2026-11-20
Long-term memory generation "automatically extracts insights" and performs "Extraction" and "Consolidation"; records are read with RetrieveMemoryRecords, which "performs a semantic search to find memory records that are most relevant to the query".
Verified 2026-11-20