Generative Agents: Interactive Simulacra of Human Behavior

The authors built a 2D sandbox world called Smallville, inspired by The Sims. It contains houses, a café, a bar, a park, a school, a dorm, and stores — each with sub-areas and interactive objects (a kitchen has a stove, a bedroom has a bed). Twenty-five unique agents inhabit this world. Each agent starts with a single paragraph of natural language backstory: who they are, what they do, and who they know.

Agents move around, enter buildings, and interact with objects. When two agents are near each other, they may start a conversation in full natural language. Users can also enter the world, observe agents, or intervene by speaking to them or changing the state of objects (e.g., setting a stove to "burning" to see if an agent reacts).

The world is represented as a tree structure: the root is the entire village, branches are areas (houses, shops), and leaves are objects (tables, shelves). Agents maintain their own partial view of this tree, updating it as they explore. This grounding lets the language model reason about spatial navigation using natural language descriptions of the environment.

The core challenge is this: a language model can generate plausible behavior for a single moment, but it cannot maintain coherence across hours, days, and evolving social relationships. Telling an agent "you are a pharmacist" and asking "what do you do next?" works once, but asking again and again leads to repetition, contradiction, and nonsense. The agent has no history, no growth, no sense of time.

The authors solve this with a three-part architecture that wraps around the language model like a cognitive scaffold. Think of it as giving the language model three tools it does not naturally have: a diary (memory stream), a journal for insights (reflection), and a calendar (planning). Each component feeds into the others, and all of them are stored as natural language — so the language model can reason about its own memories.

The memory stream is a chronological log of everything the agent has experienced. Each entry is a natural language description timestamped with its creation time and last access time. Entries include observations ("Isabella is setting out the pastries"), conversations ("Isabella and Maria are discussing the Valentine's Day party"), and even perceptions of object states ("The refrigerator is empty").

But logging everything creates a new problem: there is too much to fit into a prompt. Summarizing all memories produces generic, uninformative responses. The solution is selective retrieval: not all memories matter equally at any given moment.

Not all memories are equally useful at any moment. To decide which memories to retrieve, the architecture scores each memory using three factors.

Recency gives higher scores to recently accessed memories. It uses an exponential decay function with decay rate 0.995 per simulated game hour. Breakfast this morning matters more than breakfast last Tuesday.

Importance distinguishes mundane events from significant ones. The language model itself assigns an integer score from 1 to 10 — brushing teeth is a 1, asking your crush on a date is an 8. This score is generated once when the memory is created.

Relevance measures how related a memory is to the current situation. Each memory's text is converted into an embedding vector, and relevance is computed as the cosine similarity between that vector and a query describing the current context.

The final retrieval score is a weighted sum of these three components, each normalized to the range $[0, 1]$:

Raw observations alone are not enough. If you ask Klaus "Who would you want to spend time with?", an agent with only observations picks Wolfgang — his dorm neighbor — simply because they bump into each other most. But Klaus and Wolfgang only ever exchange pleasantries; they have no deep connection.

A better answer requires generalization: noticing that Klaus spends hours on research, and that Maria does too (in a different field), and concluding they share a passion. This is what the reflection module provides.

Reflections are triggered when accumulated importance scores of recent events exceed a threshold (150 in the paper). In practice, agents reflect roughly two or three times per simulated day. The process has two steps:

First, the model generates high-level questions from the 100 most recent memories: "What is Klaus passionate about?" "What is his relationship with Maria?" Second, it retrieves relevant memories for each question and synthesizes insights, citing evidence: "Klaus is dedicated to his research on gentrification (because of memories 1, 2, 8, 15)."

Crucially, reflections are stored back into the memory stream — so the agent can reflect on its own reflections, building a tree of increasingly abstract understanding. Leaf nodes are raw observations; inner nodes are reflections; higher nodes are reflections of reflections.

Without planning, a language model optimizes for the most plausible next action — which might mean eating lunch at 12:00, then again at 12:30, and again at 1:00. Each action is locally reasonable but globally absurd. Planning solves this by giving the agent a coherent daily schedule.

The planning process is top-down and recursive. First, the agent generates a broad daily plan in 5–8 chunks based on its persona and yesterday's summary: "1) Wake up at 8:00 AM, 2) Classes at 10:00 AM, ... 5) Work on composition from 1:00 PM to 5:00 PM." Then each chunk is decomposed into hour-long blocks, and each block into 5–15 minute actions: "4:00 PM: grab a light snack. 4:05 PM: take a short walk. 4:50 PM: clean up workspace."

Plans are stored in the memory stream too, so the retrieval function can consider them alongside observations and reflections. This means an agent deciding what to do next weighs its plan, its recent experiences, and its insights together.

Agents can also react to unexpected events. At each time step, the agent perceives its surroundings. If something noteworthy happens — a family member appears, a stove starts burning — the agent may abandon its plan and respond. The reaction is generated by prompting the language model with the observation and relevant memories, asking "Should you react, and if so, how?"

The most striking results are the social behaviors that emerge without being programmed. The authors seeded two pieces of information: Sam wants to run for mayor, and Isabella wants to throw a Valentine's Day party. Over two simulated days, three categories of emergent behavior appeared.

Information diffusion. Sam told one person about his candidacy; that person told another. By the end of two days, 8 out of 25 agents (32%) knew about the election. Similarly, Isabella's party information spread to 13 agents (52%). None of this was scripted — agents chose to share news during natural conversations.

Relationship formation. Agents who had never met introduced themselves and remembered each other later. The network density (a measure of how connected the social graph is) increased from 0.167 to 0.74 over two days. When Sam ran into Latoya at the park, she mentioned her photography project. Later, Sam remembered and asked "How is your project going?"

Coordination. Isabella invited friends to the party, spent the afternoon decorating, and enlisted Maria's help. Maria invited her crush Klaus. On Valentine's Day, five agents showed up at the right place at the right time. One agent even asked another on a date to the party — all from a single seed instruction.

The authors evaluated believability through "interviews" — asking each agent questions about self-knowledge, memory, plans, reactions, and reflections. 100 human evaluators ranked five conditions: the full architecture, three ablations (removing memory, planning, or reflection), and human-authored baselines.

Using TrueSkill rating (a generalization of the Elo chess system), the full architecture scored highest ($\mu = 29.89$). Each ablation reduced performance: no reflection ($\mu = 26.88$), no reflection and no planning ($\mu = 25.64$), and no memory, planning, or reflection ($\mu = 21.21$). The human crowdworker baseline scored $\mu = 22.95$ — below the full architecture by a large margin. The effect size between the baseline (no memory/planning/reflection) and the full architecture was $d = 8.16$ — eight standard deviations.

The ablation results show that every component contributes. Without reflection, agents cannot synthesize their experiences. Without planning, they lose temporal coherence. Without memory, they become generic language model outputs with no personal history.

The agents were not perfect. Several failure modes appeared during the simulation.

Memory retrieval failures. Agents sometimes failed to retrieve the right memory. When asked about the Valentine's Day party, one agent remembered he was supposed to discuss politics at the party, but could not remember the party existed.

Embellishment. Agents rarely fabricated experiences entirely, but they did embellish. Isabella knew Sam was running for mayor, but added that "he's going to make an announcement tomorrow" — something never discussed. Another agent described a neighbor named "Adam Smith" as having "authored Wealth of Nations," confusing him with the 18th-century economist.

Location confusion. As agents learned about more locations, they sometimes chose odd places for their activities — going to a bar for lunch when a café would be more appropriate.

Over-cooperative behavior. Due to instruction tuning in the underlying language model, agents rarely said no. Isabella accepted every party suggestion from others — Shakespearean readings, networking events — even when they conflicted with her personality.

Overly formal dialogue. Spouses greeted each other with "It was good talking to you as always" — polite but not how married couples typically speak.