Abstract:
Generative agents can simulate believable human behavior. These agents remember and reflect days past as they plan for next days.
This architecture extends LLM to store the complete record of the agents’ experiences and retrieve them dynamically to plan behavior.
observation, planning, and reflection: three parts, each contribute critically to the believability of agent behavior.
Introduction:
memory stream – a long-term memory module that records, in natural language, a comprehensive list of the agent’s experiences
reflection – synthesizes memories into higher- level inferences over time, enabling the agent to draw conclusions about itself and others to better guide its behavior.
planning – translates those conclusions and the current en- vironment into high-level action plans and then recursively into detailed behaviors for action and reaction
We manifest a small society of twenty five agents in a game environment. End users can observe and interact with these agents.
e.g. One agent wants to hold a valentine’s party. Then the system does not need to tell each agent to come to this party but this agent will automatically spread this information.
Method:
- instantiate them as characters in a simple sandbox world reminiscent of The Sims
- For each role, write a background introduction for it.
- Agents communicate with each other in full natural language. Agents are aware of other agents in their local area, and the generative agent architecture determines whether they walk by or engage in conversation.
- Users and agents can influence the state of the objects in this world, much like in sandbox games such as The Sims. For example, a bed can be occupied when an agent is sleeping, and a refrigerator can be empty when an agent uses up the ingredients to make breakfast
New discoveries – Emergent Social Behaviors
- Information diffusion: Sam tells another agent he wants to chase for mayor
- reflection: after reflection, architecture can make more believable decisions
- future directions: create more powerful simulations of human behavior to test and prototype social systems and theories, and to create new interactive experiences
Qian's Blog 