What changed
Amazon Bedrock has introduced Agentic Retrieve, a feature that allows AI agents to maintain memory of user interactions. This capability is part of Amazon Bedrock AgentCore and integrates with the Amazon Bedrock Managed Knowledge Base. Users can now configure their applications to store and retrieve conversation history, extract relevant facts, and use them to compose sub-queries for the knowledge base. This is achieved by adding a memoryConfiguration to the AgenticRetrieveStream request, which makes AgentCore memory a retrieval source alongside the knowledge base.
What this means in practice
Applications using Amazon Bedrock can now build memory-aware AI agents that remember user interactions and context. This is particularly useful for applications like customer support assistants, internal employee assistants, and document chatbots within SaaS products. The integration allows the service to restore conversation history, extract facts, and use them to generate answers grounded in the documents users are allowed to read. The user's identity is derived from the Amazon Cognito ID token, which supplies the sub (user ID) for memory and the verified email for document access control lists (ACLs). The solution ensures that each user's memory remains private and isolated, even when multiple users share the same knowledge base and memory resource.
Limitations and availability
The integration is opt-in and backward compatible, meaning existing applications can continue to function without changes. The feature is available for users who have deployed applications using the Amazon Bedrock Managed Knowledge Base. The solution overview includes an example of an internal employee assistant for AnyCompany, demonstrating how to read the user's identity from the Cognito ID token, pass it to AgenticRetrieveStream for document ACLs and memory, and design memory namespaces for multiple users. The feature is not available for applications that have not deployed the Amazon Bedrock Managed Knowledge Base.
