Amazon Bedrock Managed Knowledge Base can now combine document retrieval with AgentCore memory inside one AgenticRetrieveStream request. The integration restores conversation history, searches long-term facts when useful, generates a cited answer and saves the new exchange without a separate application database.
The request carries both a memory configuration and user context. Long-term records can include conversation summaries and explicit preferences, while access-control lists filter the knowledge base to documents that person may read. AWS recommends deriving the stable memory identity and document identity from a verified login token rather than accepting either value from the browser.
A sample employee assistant uses one shared memory resource but separates each user through namespaces and IAM conditions. Engineers can access technical runbooks, sales staff can access commission documents, and neither should retrieve the other group’s files. Trace events show whether the planner restored a session, searched memory or continued after a memory failure.
Memory is best-effort and eventually consistent. AWS observed a long-term record appearing 81 seconds after a conversation, but does not guarantee that timing. The service also adds latency and usage charges, while extracted records require an explicit deletion workflow after raw events expire. The integration removes custom history plumbing, but applications still own authentication, retention choices, permission design and monitoring for poisoned or outdated memories.