As teams move AI agents from pilot to production, a common challenge emerges. Agents can complete tasks in front of them, but they do not naturally carry learning across sessions. Without memory, corrections, preferences, and successful patterns are lost. Teams often compensate with longer prompts and repeated instructions, which increases complexity and reduces maintainability over time.

This session examines why memory is not an optional addition but a core component of production-grade agent systems. It starts with a fundamental question: if an agent is expected to improve over time, where does that learning reside? The talk explores how stateless models create friction in real-world systems and why prompting alone cannot support continuity, adaptation, and learning across sessions.

Using a document extraction use case, the session demonstrates how memory enables agents to retain useful context, adapt across runs, and improve over time. It walks through a practical loop of recall, act, evaluate, reflect, and store, showing how repeated cycles lead to more reliable outcomes. It also explains the different roles of memory types, helping teams decide what to retain, update, or discard.

What You Will Learn

• Why memory is essential for building reliable, production-ready AI agents
• How stateless models and prompt-based approaches limit continuity and learning
• How to design memory systems that support adaptation, improvement, and maintainability

Who Should Attend

• Software developers building AI agent systems
• AI and machine learning engineers
• Platform and infrastructure engineers
• Software architects
• Teams moving AI systems from pilot to production