Amazon Bedrock announced on 16 November 2026 that its AgenticRetrieveStream feature will enable memory‑aware AI agents to pull long‑term conversation history into a single API call. The update is available in the us‑west‑2 region and is designed for applications that need to remember user preferences and past interactions.
The feature works with Bedrock’s Managed Knowledge Base, which restores conversation history from short‑term memory at the start of each user interaction. Facts extracted from the dialogue are used to inform search queries within the knowledge base, allowing the agent to provide contextually relevant responses.
Privacy and Access Control
User identity is derived from the Cognito ID token, with the 'sub' claim used as a stable actorId for memory and the email claim used for document access control. Each user’s memory is isolated via a namespace template that includes the actorId, ensuring that only the intended user can see their own data.
Document visibility is filtered by a userContext that references email‑based ACLs stored in an S3 bucket. The memory resource must be in the same AWS account as the knowledge base, in the ACTIVE state, and located in us‑west‑2.
Memory Extraction Strategies
Bedrock supports both summary and preference‑based extraction strategies. Summaries capture incidental mentions, while preferences capture user‑defined settings. Long‑term memory is extracted asynchronously using the chosen strategy, and each conversation turn is saved with persistence mode set to DEFAULT.
The backend assumes an IAM role per user, with a session tag that binds memory operations to the specific user identity. The service saves each conversation turn to memory, enabling asynchronous long‑term record extraction.
In a sample implementation, 11 documents were scanned and indexed in an S3 bucket with zero failures. Three test users – two engineers and one salesperson – authenticated via Cognito and passed their identity claims to AgenticRetrieveStream. The sample required Boto3 version 1.43.72 and Python 3.10 or later.