Generative Agents

Generative agents are autonomous characters powered by large language models that maintain persistent memory stores and use them to plan future actions, react to unexpected events, and behave consistently over long interactions.

Each agent stores observations, reflections, and plans in a memory stream that is dynamically retrieved and synthesized when the agent needs to decide what to do next. The architecture separates day-to-day planning from longer-term disposition, allowing agents to balance immediate goals with stable personality traits. When multiple agents interact, their conversations are grounded in their individual memories, creating emergent social behaviors like cooperation, competition, and relationship maintenance that arise naturally rather than being explicitly programmed.

Simulating many agents at once requires substantial compute, and incoherent memory recall can break believability. The quality of the language model sets a ceiling on how natural and consistent agent behavior can be. Agents can sustain long-horizon interactions with consistent personalities, but their behavior is limited by the quality of memory retrieval and the model's ability to reason over multiple agents simultaneously.

Whether generative agents can scale to support dozens of simultaneous agents without degradation in behavior quality is unclear. How to ensure memory retrieval remains relevant over very long interaction histories is an open engineering problem. Whether emergent social behaviors are sufficiently controllable for applications requiring strict safety guarantees is an unsolved research question.