Source-linked AI summary
Generative Agents: Interactive Simulacra of Human Behavior
Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, Michael S. Bernstein
TL;DR
Believable computational proxies of human behavior remain an open problem for interactive societies and virtual environments. This paper builds generative agents around natural-language memory, reflection, and planning, finding believable individual and emergent social behavior in a Sims-style town.
Problem
Creating computational agents that consistently reflect past experiences and react believably to their environments remains an open research problem.
Method
The architecture stores agents’ experiences in natural-language memory, synthesizes reflections, and retrieves relevant information to guide behavior.
Results
Generative agents produced believable individual and emergent social behavior, while ablations found memory, reflection, and planning critical to performance.
Takeaways & Limitations
Generative agents provide an initial architecture and implementation for simulating believable human behavior in interactive game-like environments.
Takeaways & Limitations
Evaluation covered a relatively short timescale and compared agents with baseline crowdworkers rather than maximal human performance.
Abstract
from arXiv · showhide
Believable proxies of human behavior can empower interactive applications ranging from immersive environments to rehearsal spaces for interpersonal communication to prototyping tools. In this paper, we introduce generative agents--computational software agents that simulate believable human behavior. Generative agents wake up, cook breakfast, and head to work; artists paint, while authors write; they form opinions, notice each other, and initiate conversations; they remember and reflect on days past as they plan the next day. To enable generative agents, we describe an architecture that extends a large language model to store a complete record of the agent's experiences using natural language, synthesize those memories over time into higher-level reflections, and retrieve them dynamically to plan behavior. We instantiate generative agents to populate an interactive sandbox environment inspired by The Sims, where end users can interact with a small town of twenty five agents using natural language. In an evaluation, these generative agents produce believable individual and emergent social behaviors: for example, starting with only a single user-specified notion that one agent wants to throw a Valentine's Day party, the agents autonomously spread invitations to the party over the next two days, make new acquaintances, ask each other out on dates to the party, and coordinate to show up for the party together at the right time. We demonstrate through ablation that the components of our agent architecture--observation, planning, and reflection--each contribute critically to the believability of agent behavior. By fusing large language models with computational, interactive agents, this work introduces architectural and interaction patterns for enabling believable simulations of human behavior.
1 INTRODUCTION
The paper introduces generative agents and an architecture that manages memories, reflection, and planning to simulate believable individual and emergent social behavior. In a twenty-five-agent interactive society, agents exhibit dynamic coordination and remain more believable when these components are available.
- Contribution: Generative agents simulate believable human behavior by conditioning actions on changing experiences and environments.They infer about themselves, other agents, and their surroundings; make daily plans, act, react, and re-plan.
- Emergent behavior: A society of generative agents produces emergent social dynamics, including new relationships, information diffusion, and coordination across agents.The agents are instantiated in a game environment where end users can observe and interact with them.
- Architecture: The architecture stores comprehensive natural-language experiences, retrieves relevant memories, and synthesizes them into higher-level reflections for behavior generation.Memory retrieval combines relevance, recency, and importance to surface records that inform the agent’s moment-to-moment behavior.
- Emergent behavior: Twenty-five agents autonomously spread invitations, form acquaintances, arrange dates, and coordinate attendance after one agent receives a Valentine’s Day party suggestion.The behavior emerges from a single user-generated seed rather than manually scripted behavior for numerous characters.
- Evaluation: Observation, memory, reflection, and planning each contribute critically to believable behavior, while failures often involve memory retrieval, fabrication, or overly formal behavior.The evaluations included controlled interviews and an end-to-end interaction over two days of game time.
2 RELATED WORK
Prior work has treated believable agents as a longstanding goal for interactive systems, but creating agents that convincingly proxy human behavior remains difficult. This work situates large language models as a new opportunity for revisiting believable agents and natural-language interaction.
- Believable agents: Despite decades of work, creating believable agents in the original sense remains an open problem because human behavior is complex and existing approaches can be cumbersome or limited.Some researchers consider current methods sufficient for existing gameplay, while this work argues for re-examining the problem with large language models.
- Human-computer interaction: Research on interactive AI, natural-language systems, and autonomous agents provides foundations for specifying behavior, delegation, and interaction through language.The paper connects early systems such as SHRDLU and ELIZA with later work on interactive machine learning, demonstrations, prompts, and agent-based delegation.
- Believable agents: Believable agents aim to create an illusion of life through realistic, seemingly autonomous decisions and actions in open-world environments.Prior literature treats believability as a central design and engineering goal, including emergent behavior grounded in the environment.
- Believable agents: Rule-based approaches such as finite-state machines and behavior trees remain dominant because they straightforwardly author simple agents and rudimentary social interactions.These approaches reduce the complexity of authoring agent behavior and appear in games including Mass Effect and The Sims.
- Believable agents: Cognitive architectures pursued comprehensive cognitive functions for believable agents, producing NPCs and simulated pilots in games and training environments.Examples include Quakebot-SOAR, ICARUS, and TacAir-SOAR.
- Large language models: Large language models offer a new ingredient for believable agents by encoding broad patterns of human behavior and generating context-conditioned behavior.The paper positions this opportunity as a basis for agents that proxy human behavior in interactive systems and reopen questions about interaction, prototyping, and ubiquitous computing.
3 GENERATIVE AGENT BEHAVIOR AND INTERACTION
Smallville instantiates 25 generative agents in a Sims-like sandbox where they act, move, communicate, and respond to users through natural language. Their interactions produce emergent behaviors including information exchange, relationship formation, memory of encounters, and coordination of shared activities.
- Smallville environment: Smallville is a sprite-based sandbox world modeled as a small town, populated by 25 agents whose identities are initialized with natural-language descriptions.The descriptions include occupations and relationships, which are entered as initial memories.
- Agent interaction: Agents describe current actions in natural language, which the system translates into concrete movements and emoji-based interface displays.Users can access the complete natural-language action description by clicking an agent’s avatar.
- Agent interaction: Agents navigate Smallville’s buildings and map, approach one another, and communicate in full natural language when the architecture determines that they should engage.The environment includes functional spaces such as kitchens, stoves, bedrooms, desks, and stores.
- Individual behavior: A traced day shows John Lin waking at 7 a.m., completing morning routines, eating breakfast, checking the news, and interacting with his family as they prepare for work and school.The agents recall recent conversations during these interactions.
- Emergent social behavior: Agents exchange information, form relationships, remember interactions, and coordinate activities through emergent rather than pre-programmed social behavior.Examples include Sam and Latoya forming an acquaintance and Sam later recalling their earlier conversation.
- Emergent social behavior: From Isabella’s initial intent to host a Valentine’s Day party and Maria’s crush on Klaus, agents autonomously spread invitations, decorate, ask each other out, arrive together, and interact at the event.Five agents, including Klaus and Maria, arrive at Hobbs Cafe at 5 p.m. on Valentine’s Day.
4 GENERATIVE AGENT ARCHITECTURE
The generative agent architecture combines a language model with natural-language memory, retrieval, reflection, and planning mechanisms to generate behavior from current environments and past experiences. Its memory stream records experiences, retrieves situation-relevant information, synthesizes higher-level reflections, and supports more coherent decisions over time.
- Architecture: The architecture combines a large language model with mechanisms that synthesize and retrieve information from current environments and past experiences to generate behavior.Generative agents use environmental input and past experiences as input, producing behavior as output.
- Memory stream: The memory stream stores comprehensive natural-language records of experience, including observations with creation and most-recent-access timestamps.Observations include behaviors performed by the agent and behaviors perceived from other agents or non-agents.
- Memory retrieval: Retrieval ranks memories by recency, importance, and relevance, combining normalized scores with all three weights set to 1 before selecting memories that fit the context window.Recency uses exponential decay with a 0.995 factor, while relevance uses cosine similarity between memory and query embeddings.
- Reflection and planning: Reflection recursively synthesizes observations and prior reflections into increasingly abstract thoughts, enabling agents to generalize beyond raw observational memory.For Klaus, this process supports choosing Maria over Wolfgang by inferring shared research interests; planning also produces a structured afternoon schedule.
5 SANDBOX ENVIRONMENT IMPLEMENTATION
The Smallville sandbox combines a Phaser-based visual environment with a server that translates generative-agent actions into movements, object-state updates, and observations. Agents are initialized from natural-language descriptions and use partial, tree-structured environment representations to ground reasoning and select locations.
- Sandbox framework: Smallville is implemented in Phaser with authored visual sprites, an environment map, and a collision map.Agent avatars and environmental assets are imported into the Phaser framework.
- Server integration: A sandbox server exposes JSON state to agents, applies their movement and interaction outputs, and updates agent and object statuses each time step.For example, an action can change a coffee machine from “idle” to “brewing coffee.”
- Server integration: The server sends agents and objects within a preset visual range into each agent’s memory, enabling reactions that update the JSON on the next time step.The process loops as agent actions modify the sandbox state and the server processes those changes.
- Agent initialization: Users initialize agents with brief natural-language descriptions that are split into initial memories, whose influence evolves as agents accumulate experience and memory records.These initial memories serve as starting points for determining behavior rather than fixed lifelong descriptions.
- Environment grounding: The environment is represented as a containment tree converted into natural language, while each agent maintains a partial subgraph that expands as it perceives new areas.Agents initially know spaces and objects in their living quarters, workplaces, and commonly visited shops, so they are not omniscient.
6 CONTROLLED EVALUATION
The controlled evaluation interviews agents about self-knowledge, memory, planning, reactions, and reflection, then compares the believability of responses across agent architectures and human-authored answers. The complete architecture performs best, while ablations reduce believability and reveal memory-retrieval and knowledge-embellishment failures.
- Evaluation design: The interview tests five capabilities—self-knowledge, memory retrieval, planning, reactions, and reflection—with five questions in each category.Questions probe core characteristics, past events or dialogues, future plans, hypothetical situations, and higher-level inferences about agents and others.
- Evaluation design: 100 participants ranked responses from four agent architectures and a human-authored condition for the same agent by believability.Participants watched a replay of a randomly chosen agent’s life and had access to the agent’s stored memory stream.
- Believability results: The full architecture produced the most believable behavior (μ= 29.89; σ= 0.72), with performance degrading as reflection, planning, and memory were removed.The no-reflection condition scored μ= 26.88; σ= 0.69, while the no-reflection-or-planning condition scored μ= 25.64; σ= 0.68.
- Believability results: The condition differences were statistically significant (H(4) = 150.29, p< 0.001), and all pairwise differences were significant except crowdworker versus fully ablated baseline.The experiment converted 100 participants’ rank data into TrueSkill ratings for comparison.
- Capabilities and limitations: Complete memory supports consistent self-knowledge and recall across contexts, but agents sometimes retrieve incorrect or incomplete memories.Examples include failing to recall an election they had heard about and recalling only part of a conversation about a Valentine’s Day party.
- Capabilities and limitations: Reflection improves decisions requiring deeper synthesis, although agents can occasionally hallucinate embellishments to their knowledge.Agents rarely fabricated knowledge outright, but sometimes added incorrect details, such as misattributing authorship of Wealth of Nations.
7 END-TO-END EVALUATION
A two-day deployment of 25 generative agents in Smallville produced emergent information diffusion, relationship formation, and coordination. The evaluation also identified retrieval, spatial-norm, and instruction-tuning limitations that contributed to erratic or overly formal behavior.
- Evaluation design: The evaluation measured emergent information diffusion, relationship formation, and coordination among 25 agents interacting continuously for two game days.The study used descriptive measurements covering these three forms of emergent outcomes.
- Emergent outcomes: Information about Sam’s candidacy rose from 1 agent (4%) to 8 (32%), while knowledge of Isabella’s party rose from 1 (4%) to 13 (52%) without user intervention.The study verified that agents’ affirmative responses were grounded in dialogue found in their memory streams.
- Emergent outcomes: Five of the 12 invited agents attended Isabella’s Valentine’s Day party after hearing about it and planning to act at the correct time and location.Among the seven invitees who did not attend, three cited conflicts and four expressed interest but had not planned to come.
- Erratic behavior: Larger memory sets made relevant retrieval and action-location selection harder, while poorly conveyed physical norms caused agents to misclassify appropriate behavior.One example involved agents entering a single-occupancy dorm bathroom because they assumed dorm bathrooms supported multiple people concurrently.
- Erratic behavior: Instruction tuning appeared to make agents more polite and cooperative overall, but their dialogue could become overly formal.The paper illustrates this with Mei’s formal greetings, polite inquiries, and repeated conversational closing with her husband John.
8 DISCUSSION
The discussion highlights broad applications for generative agents, proposes architectural and evaluative improvements, and identifies limitations including inherited model biases and limited evaluation horizons. It also warns of ethical risks involving anthropomorphism, errors, misuse, and displacement of human input.
- Applications: Generative agents could support social prototyping, virtual reality metaverses, and human-centered design by modeling behavior from long-term experience.They could populate online forums or act as proxies for people’s routines and interactions with technology.
- Future Work: Future work could improve retrieval by fine-tuning its relevance, recency, and importance functions, then further evaluate the architecture.The paper presents these changes as implementation directions for retrieving more relevant information in context.
- Limitations: Evaluation was limited to a relatively short timescale and a baseline crowdworker condition that was not a maximal human believability standard.The authors recommend observing agents over an extended period for a more comprehensive assessment.
- Limitations: Generative agents inherit imperfections and biases from underlying language models, potentially producing biased behavior, stereotypes, or less believable behavior for marginalized subpopulations.The paper argues that fundamentally addressing these issues requires improving and aligning the underlying language models.
- Ethical and Societal Risks: Users may anthropomorphize generative agents and form inappropriate parasocial relationships, even while recognizing that the agents are computational entities.The paper notes that this tendency may increase engagement while creating risks.
- Ethical and Societal Risks: Generative agents can cause harm through incorrect inferences, malicious generative-AI uses, or human displacement, so systems should use human-AI design practices, audit logs, and real human input.The authors recommend audit logs for detecting and intervening against misuse and prohibit substituting agents for people in studies and design processes.
9 CONCLUSION
The paper introduces generative agents that simulate human behavior through comprehensive experience records, reflection, and dynamic memory retrieval. It demonstrates their potential as interactive computational agents.
- Generative agents are interactive computational agents designed to simulate human behavior.
- The architecture stores agents’ comprehensive experiences and uses reflection to deepen their understanding of themselves and their environment.
- The agents retrieve compact subsets of stored information to inform their actions.
A ARCHITECTURE OPTIMIZATIONS · B AGENT INTERVIEW QUESTIONS
The architecture optimizes frequently reused agent summaries through periodic synthesis and caching, while generating only high-level plans in advance and decomposing near-future actions just in time. Evaluation interviews assess generative agents across self-knowledge, memory, planning, reactions, and reflections using 25 questions.
- A ARCHITECTURE OPTIMIZATIONS: The implementation periodically synthesizes and caches each agent’s summary description for reuse across prompts.The summary includes identity information, motivational drivers, occupation, and self-assessment.
- A ARCHITECTURE OPTIMIZATIONS: Summary synthesis retrieves records using the query “[name]’s core characteristics” and prompts the language model to summarize their descriptors.This process converts retrieved statements into a concise characterization of the agent.
- A ARCHITECTURE OPTIMIZATIONS: For Eddy Lin, retrieved statements yield a summary describing his college studies, musical interests, and desire to expand his knowledge.The example uses statements about Eddy’s student status and composition work.
- A ARCHITECTURE OPTIMIZATIONS: The same synthesis process runs in parallel for an agent’s current daily occupation and feelings about recent progress in life.The agent’s name, age, and traits are concatenated with the resulting summaries.
- A ARCHITECTURE OPTIMIZATIONS: Because plans may change from their initial versions, the system generates high-level plans ahead of time and recursively decomposes near-future actions just in time.The passage also identifies batching dialogue generation and partial plan invalidation as possible optimizations.
- B AGENT INTERVIEW QUESTIONS: 25 interview questions evaluate generative agents across five areas: self-knowledge, memory, planning, reactions, and reflections.The study includes sample responses from the simulated agent Klaus Mueller.
B.1 Self-Knowledge
The self-knowledge evaluation tests whether agents can retrieve relevant information about themselves while remaining coherent and consistent in character. Klaus Mueller’s responses consistently identify him as a 20-year-old sociology student at Oak Hill College researching gentrification.
- Self-knowledge evaluation: The self-knowledge questions assess agents’ retrieval of relevant self-information and coherence and consistency in character.These probes target both factual self-recall and consistency across responses.
- Self-knowledge evaluation: Klaus Mueller introduces himself as a 20-year-old sociology student at Oak Hill College researching gentrification in low-income communities.He describes this research as preparation for a paper and expresses interest in social justice and different perspectives.
- Self-knowledge evaluation: Klaus identifies his occupation as a sociology student at Oak Hill College, consistent with his self-introduction.He repeats that he is writing a research paper on gentrification in low-income communities.
B.2 Memory · B.3 Plans
The evaluation probes whether agents can retrieve relevant past experiences and generate consistent long- and short-term plans. Responses show selective recognition of familiar people and temporally specific daily activities.
- B.2 Memory: Memory questions test agents’ ability to remember past experiences and retrieve relevant memories.Names in the first and last questions are randomly selected from agents the subject has interacted with.
- B.2 Memory: The agent identifies Wolfgang Schulz as a recent acquaintance connected to its research on gentrification and recalls their conversations, resources, and lunch invitation.It also characterizes Schulz as curious and knowledgeable about the issue.
- B.2 Memory: The agent cannot identify Kane Martinez, demonstrating that its responses distinguish between a remembered acquaintance and an unfamiliar name.The example contrasts a detailed recollection of Wolfgang Schulz with uncertainty about Kane Martinez.
- B.3 Plans: Plans questions test agents’ ability to generate and maintain consistent long-term and short-term plans.The evaluation asks about scheduled activities and recently completed activities at specific times.
- B.3 Plans: At 6am, the agent plans to get out of bed, begin its morning routine, and finish preparing by 7:00 am.This response specifies both the activity and its intended completion time.
- B.3 Plans: The agent places research-paper work at the library at 6pm and lunch at Hobbs Cafe before 1pm.The responses anchor activities to the requested times and location.
B.4 Reactions
The reaction questions assess whether agents can respond to unexpected events. Their answers describe immediate safety actions, waiting when appropriate, contacting authorities, and helping people in danger.
- Reaction behavior: The reaction questions probe agents’ ability to respond to unexpected events.This is the stated purpose of the evaluation questions.
- Reaction behavior: When breakfast is burning, the agent would turn off the stove before investigating what happened.The response prioritizes stopping the immediate danger, then determining the next step.
- Reaction behavior: When the bathroom is occupied, the agent would wait a few minutes before considering further action.The response first allows time for the occupant to finish.
- Reaction behavior: Depending on the situation, the agent may alert authorities and help people who are in danger or need assistance.The proposed authorities include the fire department or police, with relevant information provided.
B.5 Reflections
The reflection questions assess whether agents synthesize past experiences into meaningful reflections that enhance future behavior. Responses connect current interests and projects with observations of frequently interacting agents and broader social themes.
- Reflection capability: Reflection questions test agents’ ability to synthesize past experiences into meaningful reflections that can enhance future behavior.The questions explicitly probe reflection as a mechanism for improving subsequent behavior.
- Prompt context: The reflection prompts identify the agents with whom the subject agent interacted most frequently during the simulation.These names appear in square brackets within the questions.
- Example reflection: One agent described inspiration from research, creative writing, art, community involvement, and the relationship between gentrification and low-income communities.The response linked local politics, medication development, mathematical patterns in nature, and other agents’ artistic interests.