LLM poker agents develop theory of mind only when they can remember past games

When large language models play repeated poker games with persistent memory, they develop increasingly sophisticated models of their opponents' strategies and psychology — but only with memory enabled. Without memory, the same models never develop this social reasoning, regardless of domain knowledge or training.

This is the first evidence that something like theory of mind (modeling what another mind believes or wants) can emerge in AI purely through interaction dynamics, without explicit training. It matters because it suggests that social reasoning in AI scales with memory and interaction history, not just model size or training data. It also suggests a structural requirement: if you want AI systems that understand people, persistent memory might be as important as raw capability. The practical implication is simpler: an AI system that can't remember you won't learn to reason about you, no matter how smart it is otherwise.