SIAgentNews.com
← All news
DevelopingAgents

Amazon Bedrock adds per-user memory for private AI assistants

New features allow AI agents to remember individual user preferences and past conversations while maintaining strict data isolation.

· First reported by aws.amazon.com

Isometric diorama: picture for the story 'Amazon Bedrock adds per-user memory for private AI assistants'
AI-made isometric diorama for this story. No real people or products are shown.
30days for raw conversation event expiry

The short version

  • New memoryConfiguration allows the system to store and retrieve facts from past conversations automatically.
  • The service uses 'summary' and 'user-preference' strategies to distinguish between incidental facts and stated rules.
  • Identity-based access control ensures users only see documents they are permitted to read via ACLs.
  • IAM roles can enforce memory isolation to prevent data leakage even if application code contains errors.

Why it matters

This simplifies the development of multi-tenant AI applications by removing the need for developers to manually manage conversation history. It allows companies to build personalized assistants that remember user-specific details without risking data privacy between different employees.

The challenge of persistent context

Users expect AI assistants to remember them over time, rather than treating every interaction as a blank slate. A common frustration occurs when an application fails to recall a user's role, location, or specific preferences mentioned in a previous conversation. For example, a user might expect a policy summary to reflect their specific product tier or office location mentioned weeks ago.

Providing this level of personalization is technically difficult because the relevant context often lives outside the primary knowledge base. Developers typically have to store every conversation, extract durable facts, and decide which pieces of information are relevant to the current query. When multiple users share the same application, these steps must be perfectly isolated to ensure privacy and security.

Amazon Bedrock is addressing this by introducing Agentic Retrieve with Managed Knowledge Base and AgentCore memory. These tools take over the heavy lifting of storing conversations, extracting facts, and injecting the right information into the search process. This allows the AI to provide grounded answers that are tailored to the individual user's history.

How memory-aware retrieval works

The system functions by adding a memoryConfiguration to the AgenticRetrieveStream request. This makes AgentCore memory a retrieval source that works alongside the standard knowledge base. When a user asks a question, the service restores the conversation history from short-term memory and uses a planner to decide if a specific fact is needed.

The planner can then retrieve relevant facts from long-term memory to help compose sub-queries for the knowledge base. For instance, if a user asks about time off, the planner might pull a fact about their specific office location to refine the search. After the AI generates the final answer, the service automatically saves the new turn back into the memory resource.

To keep data private, the system uses a userContext to filter documents based on access control lists (ACLs). The knowledge base only returns results that the user is authorized to see. This ensures that while the memory resource might be shared globally, the content within it remains strictly separated by user identity.

Strategies for fact extraction

Amazon Bedrock offers two distinct strategies for how the system identifies and stores information from conversations. The summary strategy captures facts mentioned in passing, such as a user stating they are working from a specific city. These summaries help the assistant understand the general context of what has been discussed in the past.

In contrast, the user-preference strategy focuses on structured records of stated rules. This includes specific instructions like a preference for bullet points instead of tables or a requirement to always deploy to a specific region. By using both strategies, the assistant can distinguish between incidental conversation history and the explicit way a user wants to be helped.

The layout of these memories is managed through namespace paths. Developers can organize these paths so that the planner can reach both profile information and conversation summaries. Because the system uses unique actor IDs, a new user automatically gets their own private space without the need for manual provisioning for every new account.

Security and multi-tenant isolation

Security is a primary focus for enterprise applications where multiple employees share the same knowledge base. The system uses Amazon Cognito to authenticate users and provide identity tokens. These tokens contain a stable ID that identifies the user's memory and an email address that identifies them to the knowledge base's ACL filter.

Using a stable ID is critical because email addresses can change or be reassigned. By using a unique identifier that never moves between people, the system ensures that a new employee doesn't inherit the memory of a previous staff member. The application only needs to pass the user's identity; the service handles the rest of the isolation.

Furthermore, AWS Identity and Access Management (IAM) can be used to enforce memory isolation at the infrastructure level. This provides a safety net that prevents data leakage even if there is a mistake in the application code. This architecture allows a single company to host an assistant for everyone while keeping individual memories and document access private.

Implementation and next steps

The new integration is designed to be backward compatible and opt-in for existing applications. Developers who previously had to manually persist conversation history can now replace those steps with the managed memory service. This reduces the amount of custom code required to maintain a stateful experience for users.

To get started, developers need to create a managed knowledge base with an ACL-enabled data source. They must also set up a memory resource and configure the specific strategies they want to use for fact extraction. The service handles the ingestion, indexing, ranking, and planning required to make these components work together seamlessly.

As organizations move toward more agentic workflows, the ability to provide context-aware responses will become standard. Companies can now build assistants that feel more human by remembering past interactions while maintaining the strict data boundaries required by corporate security policies. This shift allows for more complex, multi-turn interactions that are grounded in both company data and personal user history.

What to watch

  • How the planner decides which specific long-term facts are relevant to a query.
  • The performance impact of speculative retrieval on large-scale knowledge bases.
  • Future expansion of memory strategies beyond summaries and preferences.

Try it

Check the Amazon Bedrock documentation for Agentic Retrieve to see the sample architecture for multi-tenant chat applications.

Sources and links

First reported by aws.amazon.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news