The paper introduces generative agents, a class of computational agents that simulate believable human behavior by extending large language models (LLMs) with a structured memory system, reflection mechanisms, and planning capabilities. The authors design an architecture comprising a memory stream that records all observations in natural language, a retrieval module that scores memories by recency, importance, and relevance, a reflection process that synthesizes higher‑level insights from recent memories, and a planning component that generates multi‑step daily agendas and reacts to unexpected events. They instantiate this architecture in a small‑scale sandbox world called Smallville, reminiscent of The Sims, populated by 25 agents each initialized with a brief persona description. Agents perceive the environment, store observations, retrieve relevant memories, reflect, plan, and act, producing natural‑language dialogues and emoji‑based action visualizations. The system supports user interaction via natural‑language commands and allows agents to influence the sandbox state. Two evaluations are conducted: (1) a controlled interview‑based study where agents are queried on self‑knowledge, memory, planning, reaction, and reflection, comparing the full architecture against three ablations (removing reflection, planning, or both) and a human‑crowdworker baseline; results show the full model achieves the highest believability scores (TrueSkill 29.89) and each component contributes significantly. (2) an end‑to‑end two‑day simulation assessing emergent social phenomena such as information diffusion, relationship formation, and coordination (e.g., a Valentine’s Day party). The study finds substantial diffusion of event information, increased network density, and partial coordination, while also identifying failure modes (memory retrieval errors, inappropriate location choices, over‑cooperativeness). The paper adds a detailed discussion of architecture optimizations—caching agent summaries, hierarchical plan decomposition, batch dialogue generation, and parallel execution—to improve runtime efficiency. It also provides the full set of 25 interview questions used for evaluation, along with sample responses from a simulated agent (Klaus Mueller) that illustrate self‑knowledge, memory recall, planning, reaction to unexpected events, and reflective reasoning. The authors discuss broader applications (design prototyping, social simulation, immersive environments), future technical improvements (better retrieval, cost reduction, parallelization), limitations, and ethical considerations (parasocial relationships, hallucinations, bias, misuse). They conclude that generative agents provide a promising foundation for interactive systems that require long‑term, human‑like behavior.
Topics
Generative Agents: Defining Believable Human‑Like Proxies
Introduces the concept of generative agents, their purpose as believable simulacra for interactive applications, and the motivation behind extending LLMs with memory and planning.
Core Architecture: Memory Stream, Retrieval, Reflection, and Planning
Details the four‑component architecture—natural‑language memory storage, scoring‑based retrieval, periodic reflection to synthesize high‑level insights, and hierarchical planning for daily actions.
Architecture Optimizations for Real‑Time Performance
Describes practical engineering improvements such as cached agent summary synthesis, hierarchical plan decomposition, batch dialogue generation, and parallel execution to reduce latency and computational cost.
Smallville Sandbox: Implementation of a Sims‑Style World
Describes the construction of the Smallville environment, agent avatars, natural‑language ↔ game‑state translation, and user interaction mechanisms.
Controlled Interview Evaluation and Sample Agent Responses
Presents the 25‑question interview protocol probing self‑knowledge, memory, planning, reactions, and reflections, and includes exemplar answers from a simulated agent (Klaus Mueller) that demonstrate the system’s capabilities.
Controlled Interview Evaluation and Ablation Study
Methodology and results comparing the full architecture against ablations (removing reflection, planning, or both) and a human baseline, reporting believability scores (TrueSkill) and component contributions.
End‑to‑End Two‑Day Simulation: Diffusion, Network Density, and Coordination
Analyzes long‑term emergent outcomes, quantifies information diffusion, relationship network growth, and party attendance, and discusses observed failure modes.
Applications, Future Directions, and Limitations
Explores potential uses in design prototyping, social computing, VR/AR, outlines technical improvements (retrieval tuning, parallel agents), and acknowledges cost and scalability constraints.