AWS re:Invent 2025 - Make agents remember with Amazon Bedrock AgentCore Memory (AIM331)

Summary of AWS re:Invent 2025 - Make agents remember with Amazon Bedrock AgentCore Memory (AIM331)

Introduction to Agent Memory

• Importance of providing the right context to agents at the right time
• Challenges with agents lacking memory, such as forgetting user preferences and conversation history
• Example of building a presentation creation agent with and without memory

Short-Term Memory

• Capturing raw interaction history as events with actor ID and session ID
• Maintaining conversation history for recent interactions
• Preserving interaction state to resume from previous sessions

Long-Term Memory

• Automatically extracting key insights from short-term memory using large language models
• Built-in memory strategies: summary, user preferences, semantic, override, and self-managed
• Ability to retrieve long-term memory records semantically based on queries

Enterprise-Scale Memory Management

• Experian's journey from product-specific short-term memory to a unified long-term memory architecture
• Limitations of short-term memory: continuity, performance, cost, cross-product recall, and compliance
• Experian's design principles: evaluation and test frameworks, cross-agent interoperability, targeted context windows, and namespace-based isolation

Demonstration of Agent Memory in Action

• Comparison of a basic agent without memory and an agent with memory-enabled capabilities
• Agent with memory able to remember user preferences for presentation styling and content
• Automatic application of user preferences to generate a tailored presentation

Key Takeaways and Next Steps

• Integrating agent memory to improve user experience and minimize repetitive instructions
• Potential use cases: coding assistants, customer support agents, and other agentic applications
• Invitation to attend Dr. Swami's keynote on new agent core product features
• Encouragement to build and share use cases with the presenters

Technical Details

• Use of Amazon Bedrock, Anthropic Claude 4.5, and Sonnet 4.5 models
• Integration with IAM permissions and security principles
• Short-term memory implementation using a database or key-value store
• Long-term memory using vector stores and large language model-based strategies

Business Impact

• Enhancing user experience and customer satisfaction by remembering preferences
• Improving agent efficiency and reducing repetitive instructions
• Enabling more personalized and contextual interactions
• Potential cost savings by reducing support calls and improving first-call resolution

Examples

• Presentation creation agent remembering user's styling preferences and content inclusions
• Coding assistant agent retaining user's preferred coding style and minimal file creation instructions
• Customer support agent recalling past issues to provide better-informed assistance