Build memory-aware AI agents with Agentic Retrieve on Amazon Bedrock Managed Knowledge Base
Amazon Bedrock AgentCore now supports memory‑aware agents via the AgenticRetrieveStream API, letting developers add a memoryConfiguration that treats long‑term user facts as a retrieval source alongside a Managed Knowledge Base. The planner automatically decides when to pull stored user facts to enrich sub‑queries, and the service writes each turn back to memory while respecting per‑user ACLs supplied via a userContext. This enables multi‑tenant assistants that keep both document access and conversational memory isolated per user without custom persistence code.
⚡ Key Takeaways
- Adding a memoryConfiguration to the AgenticRetrieveStream request enables AgentCore memory as a retrieval source.
- The planner can inject retrieved user facts into sub‑queries sent to the Managed Knowledge Base.
- Per‑user isolation is enforced by passing the Cognito ID token in userContext and configuring IAM policies for memory namespaces.
- Memory namespaces must be designed per user to keep each user’s stored facts private.
- The integration is opt‑in and backward compatible; existing applications only need to authenticate users and supply their identity.
Engineers can now build scalable, multi‑tenant AI assistants that remember users across sessions while guaranteeing data privacy and ACL‑driven document grounding, reducing custom state‑management overhead.
✅ Practical Steps
- Add a memoryConfiguration field to your AgenticRetrieveStream request to enable AgentCore memory retrieval.
- Pass the authenticated user's Cognito ID token in the userContext parameter so the knowledge base and memory respect ACLs.
- Define per‑user memory namespaces and configure IAM policies to enforce isolation, then read the trace to verify when the planner accessed memory.
Want the full story? Read the original article.
Read on AWS ML Blog ↗