Generative Agent Architecture

Condition long-horizon agent behavior on a memory stream, recursive reflection, and hierarchical plans rather than single-shot LM prompts.

Generative agent architecture (Park et al.) extends an LLM with three coupled loops so behavior stays coherent as experience grows: record perceptions in a memory stream, periodically synthesize reflection trees, and maintain hierarchical plans that decompose into actionable steps. Reflections and plans write back into the same stream and are retrieved like observations.

Plain few-shot or chain-of-thought prompts condition only on the current context window; they cannot hold days of multi-agent social history. This architecture is for believable simulacra under cascading interactions — not a claim of genuine agency — and remains the spine even as base models improve. Failure modes to design for: retrieval misses, embellished memories, and overly formal speech inherited from instruction tuning.

For harness builders, treat it as an instance of OS-inspired agent memory plus reflection and error correction at social timescales, with a perceive → store → continue-or-react → (re)plan action loop akin to ReAct but driven by retrieved biography rather than only the latest tool observation. Ground agents in an agent environment tree so location and object state are partial and updatable. Expect emergent multi-agent coordination when many such agents share a sandbox — diffusion and parties are architecture tests, not feature flags. Controlled interviews after simulation days show believability rises with each memory/reflection/planning layer; ablations without reflection fail to synthesize (e.g. gift suggestions from many meetings). End-to-end metrics: information diffusion, relationship-graph density, and coordination attendance. Design against memory hacking, embellished recalls, and instruction- tuning over-cooperativeness. For human steering, speaking as an agent’s inner voice (directive persona) is more likely obeyed than an external interlocutor; expose that as an agent UX steering mode alongside ordinary NL conversation.

---

Sources

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