AI Agents That Remember: Building Stateful Systems with Lakebase

Description

While many teams attempt to build AI agent memory through basic LLM prompts and traditional databases, they often miss 75% of what makes an agent truly intelligent. This session explores the critical role of stateful systems in agent development and introduces Lakebase, a managed serverless Postgres architecture designed to bridge the gap between OLAP and OLTP. Savannah Longoria breaks down the four distinct types of agent memory—working, episodic, entity, and procedural—and demonstrates how to overcome infrastructure complexity using Lakebase features like instant zero-copy branching and real-time data synchronization. Key Takeaways: - The Four Pillars of Agent Memory: Understanding the differences between working, episodic, entity, and procedural memory to build truly intelligent agents. - Overcoming the Stateless Limitation: Solving the scaling and reliability problems inherent in stateless LLMs by implementing external persistence layers. - Unified Infrastructure with Lakebase: Bringing managed serverless Postgres directly into the Lakehouse to eliminate fragile ETL pipelines and siloed data. - Safe Experimentation via Branching: Utilizing git-style database branching for zero-copy isolation, allowing developers to test agent changes against production data without risk.