The paper "Generative Agents: Interactive Simulacra of Human Behavior" introduces "generative agents," which are computational software agents designed to simulate believable human behavior. To demonstrate their capabilities, researchers populated a sandbox environment reminiscent of The Sims with twenty-five unique agents who can plan their days, share news, form relationships, and coordinate group activities.

To enable these agents to maintain long-term coherence and act consistently over time, the authors developed a novel architecture that connects a large language model to three key components:

• Memory and Retrieval: Agents possess a "memory stream" that comprehensively records their experiences in natural language. When deciding how to act, a retrieval model surfaces specific memories based on a combined score of their recency, importance (distinguishing mundane events from core memories), and relevance to the current situation.
• Reflection: Because raw observational memory is not enough to make generalizations, agents periodically synthesize their memories into higher-level inferences. This allows them to draw abstract conclusions about themselves and others.
• Planning and Reacting: Agents translate their conclusions and current environment into high-level daily plans, which are recursively broken down into detailed moment-to-moment behaviors. Agents can also dynamically react and alter these plans when they observe unexpected changes in their environment or converse with other agents.

When deployed over a two-day simulation, the community of generative agents exhibited emergent social behaviors without user intervention, such as autonomously spreading information about a mayoral election, forming new ties, and coordinating attendance for a Valentine's Day party. Through ablation studies, the researchers found that removing any of the core architectural components—observation, planning, or reflection—degraded performance, and the full generative agent architecture successfully produced more believable behavior than human crowdworkers.