The paper introduces a generative‑agent framework that augments large language models (LLMs) with a natural‑language memory stream, retrieval, reflection, and planning modules to produce agents that act and converse like humans. Agents store observations as textual memories, retrieve relevant memories using similarity search, reflect to infer higher‑level insights, and generate plans that are recursively decomposed into moment‑to‑moment actions. The authors implement the system using GPT‑3.5‑based LLMs, cache core characteristic summaries for efficiency, and parallelize agent execution to approach real‑time simulation. A sandbox world called Smallville contains 25 agents whose interactions generate believable individual behavior and emergent social dynamics (e.g., gossip, friendships, conflicts). Evaluation is performed through a structured interview protocol covering self‑knowledge, memory, planning, reactions, and reflections, with sample responses from an agent (Klaus Mueller) illustrating coherence and consistency. The paper situates its contribution within a broad literature on cognitive architectures, game AI, interactive narrative, and human‑AI interaction, citing works ranging from Soar and Soar‑based models to recent LLM prompting research. Ethical considerations, such as the potential for deceptive agents and the need for transparency, are discussed. Overall, the work demonstrates that extending LLMs with memory‑centric cognitive loops enables scalable, socially aware agents suitable for interactive entertainment and social computing research.
Topics
Memory‑Centric Cognitive Loop
Agents continuously record observations as natural‑language memories, retrieve relevant memories through similarity search, and run a reflection step that synthesizes higher‑level insights. This loop supplies the contextual grounding for all subsequent reasoning, planning, and dialogue generation.