Context Is Their Reality: How AI Agents Perceive the World - Elena Kolevska

An AI agent can only act on what it knows. And everything it knows about you exists only as far as it's been loaded into the context window. You get that window right and the agent is brilliant. Get it wrong, or stale, or half-empty, and its reality starts to drift from yours. This talk looks at context not only as an engineering constraint, but also as a philosophical problem. What does it mean for a machine to know something? What does it mean for it to forget? And when what the agent believes and what's actually true come apart, who's responsible for the gap? If you're building agents, the answer is: you are. The model can't handle this for you, the gap is yours to manage. So the rest of the talk gets practical: how the context window actually works, why agents are stateless by default, and what that statelessness costs you in money, quality, and trust. We finish with a concrete architecture for giving an agent durable, intelligent memory: one that holds onto past data and conversations, pulls back only what matters in the moment, and keeps the agent's reality close to your own.