Generative agents are AI-powered software agents that simulate human-like behavior, making them useful for a range of applications such as immersive environments, communication rehearsal spaces, and prototyping tools. These agents can perform daily activities, express creativity, form opinions, and even initiate conversations with others. They remember their past experiences, reflect on them, and use them to plan future behavior.

The architecture of generative agents is built upon large language models, allowing them to store a complete record of their experiences in natural language. These experiences are synthesized into higher-level reflections and retrieved dynamically to guide the agents' behavior. In a simulated sandbox environment inspired by The Sims, the generative agents demonstrated believable individual and social behaviors, such as autonomously planning and attending a Valentine's Day party. This research showcases the potential of fusing large language models with interactive agents to create realistic simulations of human behavior.

Read the original paper for more details or check out the Ars Technica article on this.