The paper introduces Generative Agents: Interactive Simulacra of Human Behavior, computational software agents designed to simulate believable human behavior in interactive environments. The core idea is to extend LLMs with an architecture that allows agents to maintain long-term consistency, remember and synthesize past experiences, and dynamically plan their actions and reactions based on their environment and history. The authors demonstrate these agents in a sandbox environment reminiscent of The Sims.
The generative agent architecture consists of three main components:
- Memory Stream: This is the agent's long-term memory, storing a comprehensive record of its experiences in natural language. The basic unit is an "observation," representing events perceived by the agent. Each memory object includes a natural language description, creation timestamp, and last access timestamp.
- Retrieval: To manage the large memory stream and fit within the limited context window of an LLM, a retrieval function selects the most relevant memories for the agent's current situation. This function scores memories based on a weighted combination of:
- Recency: More recently accessed memories score higher (exponential decay).
- Importance: Mundane vs. poignant memories are scored by prompting the LLM (1-10 scale).
- Relevance: Memories semantically related to the current situation (query) score higher, calculated using cosine similarity of embedding vectors. The top-scoring memories that fit the context window are passed to the LLM.
- Reflection: This process synthesizes observations and existing reflections into higher-level, more abstract thoughts. Reflections are generated periodically when the cumulative importance of recent events exceeds a threshold. The process involves:
- Identifying salient questions from recent memories using the LLM.
- Retrieving relevant memories (including other reflections) based on these questions.
- Prompting the LLM to generate insights based on the retrieved memories, citing the evidence. These generated reflections are added to the memory stream, forming a hierarchical structure or "reflection tree" where insights build upon lower-level memories. Reflection enables agents to generalize and make inferences that guide their behavior.
- Planning: To ensure long-term coherence, agents create hierarchical plans. These plans are stored in the memory stream and influence future actions.
- Initial high-level daily plans (5-8 chunks) are generated using the LLM based on the agent's summary description and previous day's activities.
- These high-level plans are recursively decomposed into finer-grained actions (e.g., hour-long chunks, then 5-15 minute chunks).
- Plans include a location, start time, and duration.
Reacting and Updating Plans: Generative agents operate in a continuous action loop. At each time step, they perceive their environment (observations are added to memory). The architecture prompts the LLM with recent observations and relevant context retrieved from memory to determine if the agent should react. If a reaction is triggered, the agent's current plan is regenerated starting from that point.
Dialogue: Conversations between agents are generated by conditioning the LLM on the agents' memories about each other and the dialogue history. This allows agents to converse believably based on their relationship and shared experiences.
Sandbox Environment Implementation:
The paper demonstrates generative agents in a 2D sandbox world called Smallville, built using the Phaser framework.
- The environment (areas, objects) is represented as a tree structure, translated to natural language for the agents (e.g., "there is a stove in the kitchen").
- Agents maintain their own subgraph of the environment tree based on what they have perceived.
- A server mediates interaction between the agents and the game engine, updating agent locations and object states based on agent actions.
- Agent actions are generated by prompting the LLM and recursively traversing the agent's environment tree to determine specific locations (e.g., identifying the correct room and object for an action).
- Initial agent identities are provided as a paragraph description, split into initial seed memories.
Evaluation:
The authors conducted two evaluations:
- Controlled Evaluation: Agents were "interviewed" with natural language questions across five categories: self-knowledge, memory, planning, reactions, and reflections. Human evaluators ranked the believability of responses generated by the full architecture, three ablated versions (removing memory, reflection, and/or planning), and a human crowdworker baseline. Results showed the full architecture was significantly more believable than all ablations and the crowdworker baseline, demonstrating the importance of each architectural component. Qualitative analysis revealed agents could retrieve memories but sometimes embellished them, and reflection was crucial for synthesizing experiences into deeper insights.
- End-to-End Evaluation: A simulation with 25 agents ran for two game days to observe emergent social behaviors. Measurements tracked:
- Information Diffusion: The spread of knowledge about Sam's mayoral candidacy and Isabella's party was measured via interviews. Information spread successfully among agents.
- Relationship Formation: Network density of mutual knowledge increased significantly, indicating agents formed new relationships.
- Coordination: Agents successfully coordinated to attend Isabella's Valentine's Day party after being invited. The simulation also revealed boundary conditions and errors, such as agents choosing less typical locations with larger memory, misinterpreting physical norms of the environment, and exhibiting overly polite or cooperative behavior potentially due to LLM instruction tuning.
Applications and Future Work:
The paper suggests generative agents could be applied in social prototyping (populating online forums or virtual worlds), training simulations, human-centered design proxies (modeling user behavior for system design), and creating more dynamic game NPCs. Limitations and future work include improving the memory retrieval function, optimizing computational performance and cost, conducting longer-term and more rigorous evaluations, testing robustness to prompt/memory hacking, and addressing inherent biases and limitations inherited from the underlying LLMs.
Ethics and Societal Impact:
The authors discuss important ethical risks:
- Parasocial relationships: Users may anthropomorphize agents, leading to over-reliance or emotional attachment. Mitigations proposed are explicit disclosure of computational nature and value-alignment to prevent inappropriate reciprocity.
- Impact of errors: Errors in agent inference could cause harm in applications like ubiquitous computing. Focusing on lower-stakes environments like games and following human-AI design best practices are suggested.
- Exacerbating Generative AI risks: Agents could contribute to deepfakes, misinformation, or tailored persuasion. Maintaining audit logs is proposed as a deterrent and detection mechanism.
- Over-reliance: Developers might use agents to displace human input in design processes. The authors stress agents should complement, not replace, real human stakeholders.