Agent Environment Tree

Represent places and objects as a containment tree rendered to natural language, with each agent holding only a personal subgraph that can go stale.

Long-horizon sandbox agents need a world model richer than the latest observation string. An environment tree stores areas and objects with containment edges (“stove in the kitchen”), rendered to NL for prompting. Each agent keeps a personal subgraph of what it has seen (home, workplace, common shops), updated on navigation — not an omniscient shared map. Subtrees can go stale until re-entry.

Action location selection walks the agent’s tree from root to a leaf via recursive LM prompts (prefer staying put when the activity fits the current area), then pathfinding animates movement. Object state updates similarly: query the LM to map an NL action onto status changes (espresso → machine brewing). Users or other agents can rewrite object status in NL (“stove is burning”) so the next perceive → store cycle notices and reacts — a concrete affordance for generative agent architecture and world-model-augmented planning.

---

Sources

• Generative Agents (Park et al.), pp. 1–22