#Generative Agents

In 2023, the 'Generative Agents' paper from Stanford and Google introduced a way to create believable digital characters that can plan their days, form relationships, and coordinate activities autonomously. While previous non-player characters (NPCs) in games relied on rigid scripts or simple state machines, these agents used large language models to simulate the complexity of human life. The researchers populated a sandbox world with 25 agents and observed how individual actions coalesced into social dynamics. It was a shift from programming behaviors to architecting memories.

#The Ranked Memory Stream

The generative agent architecture: perceiving, remembering, reflecting, and planning in a continuous loop.

The primary technical shift in Generative Agents was an architecture that prioritized the storage and retrieval of long-term experience via a 'Memory Stream.' Because the context window of a language model is finite, the researchers developed a ranking function to decide which memories to inject into the agent's current prompt. This function weights three specific components: recency, importance, and relevance. Recency uses an exponential decay to prioritize fresh observations, while importance is a score (1–10) generated by the model itself to distinguish mundane tasks from significant life events. Relevance is calculated using cosine similarity between the current situation and historical text. This system revealed that believable behavior is not a result of complex logic, but of a model's ability to recall the right information at the right time.