Remember Me? The Engineering Decisions Behind AI Agent Memory
About this session
We started with what sounded like a simple request: "Our agents should remember users." That turned out to be a much more complicated problem than we expected…
What deserves to become a memory? Should you keep entire conversations or extract facts? What about insights that are never explicitly stated by the user? What happens when today's memory contradicts yesterday's? When should an agent use memory, and when is it better to ignore it? And why does every team somehow end up building a different memory solution?
Using examples from building a shared memory layer for AI agents at Intuit, we'll explore what we chose to remember, how memories evolve over time, how they're retrieved at the right moment, and what we'd do differently today. You'll leave with a practical framework for designing agent memory, architectural tradeoffs we discovered along the way, production lessons, and hopefully fewer existential debates with your own agents.
Speaker
Key takeaways
- practical framework for designing memory in AI agents - audience can apply immediately to their own systems
- understand what deserves to be remembered, how memories evolve and how to retrieve them effectively
- learn where to draw the line between a shared memory platform and individual agents