Underlying paper - Generative Agents

Simulates autonomous agent behavior by combining memory retrieval (storing and recalling past interactions), planning (decomposing goals into sub-tasks), and action execution. Agents maintain a persistent memory stream of observations and interactions, retrieve relevant memories based on current context, and use retrieved memories to inform planning and decision-making. The architecture uses a hierarchical action planning system where high-level goals are decomposed into concrete actions, with memory-informed reasoning at each step.

Retrieves relevant memories from an agent's memory stream using a combination of semantic similarity (embedding-based matching) and temporal/relevance weighting. The system scores memories based on how semantically similar they are to the current query context, then re-ranks by recency and importance. This enables agents to surface the most contextually appropriate past experiences when making decisions, without requiring explicit memory management or manual tagging.

Simulates how information spreads through the agent population via natural dialogue and interaction. When agents interact and exchange information, the system tracks what information each agent knows and updates their knowledge based on conversations. This enables emergent information propagation where rumors, news, and knowledge spread through the agent network based on who talks to whom, creating realistic social dynamics where information availability varies across agents.

Decomposes high-level agent goals into concrete, time-aware sub-tasks and actions through a multi-step planning process. Given a goal (e.g., 'attend a party'), the system generates intermediate steps (e.g., 'get dressed', 'walk to location'), then grounds each step into specific actions with estimated durations. The planner uses memory-retrieved context about the agent's current state, environment, and past experiences to make planning decisions, ensuring generated actions are feasible and contextually appropriate.

Generates realistic interactions between agents by using language models to synthesize dialogue and reactions based on each agent's memory, personality, and current goals. When two agents interact, the system retrieves relevant memories for each agent, constructs a prompt that includes both agents' context and the interaction scenario, and generates dialogue and actions that reflect each agent's perspective. The generated interactions are then added to both agents' memory streams, creating a shared interaction history.

Maintains a chronological log of all observations, interactions, and thoughts for each agent, stored as a time-indexed memory stream. As agents act and perceive their environment, new memories are automatically added to the stream with timestamps and metadata (type: observation/interaction/thought, importance level, involved parties). The memory stream serves as the agent's persistent state and ground truth for what has happened, enabling agents to maintain continuity across simulation steps and retrieve context for decision-making.

Generates observations of the environment and other agents by querying the current simulation state and converting it into natural language descriptions that agents can perceive. When an agent is in a location, the system generates descriptions of what the agent observes (other agents present, objects, activities), formatted as natural language observations that are added to the agent's memory stream. This enables agents to perceive their environment without explicit sensor models, using language as the interface between the simulation state and agent cognition.

Initializes agents with a personality profile, initial goals, and background context that shapes their behavior throughout the simulation. Each agent is created with a name, age, personality traits, relationships with other agents, and initial goals. This initialization context is stored in the agent's memory stream and used to condition all subsequent reasoning, planning, and interaction generation, ensuring agents maintain consistent personality and motivation throughout the simulation.

Manages simulation time progression by stepping through discrete time intervals and scheduling agent actions based on their planned action sequences and estimated durations. The system maintains a global simulation clock, tracks when each agent's current action will complete, and triggers the next planning cycle when an agent finishes their current action. This enables realistic temporal dynamics where agents are occupied with actions for varying durations and interact asynchronously based on their schedules.

Maintains the shared simulation environment state including agent locations, object positions, and environmental properties. The system tracks which agents are in which locations, what objects are present, and any environmental state changes resulting from agent actions. This centralized state representation enables agents to perceive their environment, interact with objects and other agents, and have their actions affect the shared world state in a consistent manner.

Enables agents to generate internal thoughts and reflections about their experiences, goals, and observations by prompting the language model to generate agent thoughts conditioned on recent memories and current context. These generated thoughts are added to the agent's memory stream, creating a record of the agent's internal reasoning process. This enables agents to consolidate experiences, plan future actions, and maintain coherent long-term goals through explicit reflection.