Architecture of Generative Agents
Last updated
Last updated
Generative Agents are AI-driven characters built on large language models (LLMs) that simulate realistic human-like behavior. They are designed for open-world environments, where they can interact with other agents, respond to changing contexts, and autonomously evolve through memory, reasoning, and planning.
At the heart of this architecture is a novel agent framework that integrates LLMs with memory retrieval and synthesis mechanisms. This ensures that agents generate behavior not just from prompt inputs, but from a coherent and personalized internal history.
Without such mechanisms, an LLM might generate plausible actions in isolation — but fail to exhibit continuity, react to past events, or demonstrate consistent identity. The key challenge, therefore, is enabling agents to retrieve, reflect on, and reason over their memory in a way that informs long-term, believable behavior.This architecture is structured around three core components:
Memory Stream & Retrieval
Reflection
Planning & Action