AWS re:Invent 2025 - Put your data to work for Agentic AI with AWS storage (STG218)

Putting Data to Work for Agentic AI with AWS Storage

The Importance of Agentic AI

• Agentic AI systems can autonomously achieve business outcomes, optimize operations, and make real-time decisions
• However, most organizations have vast amounts of data that their AI agents cannot effectively use
• The breakthrough lies in transforming passive storage into active memory that agents can access and learn from

Key Characteristics of Agentic AI

• Autonomy: Agentic AI systems don't just respond to prompts, but actively work towards specific goals
• Reasoning Capabilities: Agentic AI leverages large language models (LLMs) for understanding and decision-making
• Tools: Agents can interact with internal and external systems to execute actions
• Memory: Agents maintain continuity and build knowledge over time
• Context Awareness: Agents adapt their behavior based on the environment and who they are interacting with
• Prompt Engineering: Defines the agent's role, capabilities, and constraints

The Importance of Agent Memory

• Agent memory is a computational "exocortex" that combines LLM memory with persistent storage
• This allows agents to remember, retrieve, and adapt over time based on past experiences and new information
• Key capabilities enabled by agent memory:
  • Contextual Intelligence: Agents understand the "why" behind user requests
  • User Preferences: Agents personalize interactions based on individual communication styles and behavioral patterns
  • Knowledge Retention: Agents build their own knowledge base and get smarter with each interaction

Types of Agent Memory

1. Short-Term Memory:

   • Maintains conversation flow and immediate context
   • Requires fast, low-latency access to recent messages and state
   • Includes working memory (actively processed information) and broader short-term memory

2. Episodic Memory:

   • Agent's record of specific events and interactions
   • Includes conversation history, summaries, and metadata

3. Semantic Memory:

   • Agent's organized knowledge base of facts, concepts, and relationships
   • Enables consistent reasoning and decision-making

4. Summary Memory:

   • Distills key insights from longer interactions to enable scalable retrieval

Building Scalable Agentic AI

• Open-source frameworks like Hugging Face Transformers, LangChain, and Llama Agent can accelerate experimentation
• Challenges in moving to production include scaling infrastructure, managing security/governance, and integrating various memory components
• Managed services like Amazon Bedrock can simplify the process by handling infrastructure, memory management, and orchestration

The Role of AWS Storage

1. Data Lake Foundation:

   • S3 as the foundation, with services like S3 Tables and Iceberg for data organization and discoverability
   • Enables agents to access structured and unstructured data across the enterprise

2. Short-Term Memory:

   • FSx file services for low-latency, scalable shared memory access
   • DynamoDB for highly transactional state management
   • Elasticache for semantic caching to reduce LLM interactions

3. Long-Term Memory:

   • S3 Vectors for scalable, cost-effective semantic search and similarity matching
   • Enables agents to build knowledge bases and personalized user profiles over time

4. Integration and Discoverability:

   • Model Context Protocol (MCP) for standardized agent-to-service communication
   • S3 Metadata for enabling agents to discover and understand available data

Real-World Examples

• Rocket Companies used Amazon Bedrock Agents to build an agent-powered engagement platform for their customers, improving query resolution and customer satisfaction

Key Takeaways

1. Build a modern, iceberg-based data foundation on S3 to make data actionable for agents
2. Use S3 Vectors and other AWS storage services to scale agent memory and semantic search cost-effectively
3. Implement observability and rapid iteration to continuously improve agent accuracy and business value
4. Consider managed services like Amazon Bedrock to reduce undifferentiated heavy lifting and accelerate agent deployment