Most AI systems today operate in stateless environments, relying on limited prompt windows to infer context. But what happens when context isn’t just a feature, it’s the lifeblood? In this talk, Brian Walters explores the design and implementation of scalable, long-term memory for AI systems using TiDB. Drawing from real-world architectures and use cases, he walks through how TiDB’s hybrid transactional and analytical engine enables persistent memory that is queryable, versioned, and multi-modal.