From Rule-Based Systems to LLM Agents: A Survey of NPC Architectures, Memory Systems, and Integration Challenges

NPCs play a vital role in modern video games, making game engaging, story progress possible, and the development of a convincing virtual world. In response to an increase in games with an open world architecture, there is an increased demand for responsive and believable NPCs. The rule-based approach allows for predictable and controllable behavior yet is inflexible. Machine learning approaches allow for responsive behavior; however, such behavior is usually hard to explain and make fit into the story. Recent progress in large language models (LLMs) allows one to generate contextual dialogue and develop more realistic interaction with NPCs. In this paper, we conduct a survey on the evolution of NPCs through three different architectures: rule-based architecture, machine learning-based approaches, and LLM-based agents, paying special attention to the problem of memory structures, decision-making systems, and problems of integration. It is found that none of the approaches provides the desired features for adaptability, explainability, story integration, and behavioral coherence. Architecture fragmentation or the absence of strong links between memory, reasoning, dialogue generation, and decision-making becomes the primary problem. To solve this problem, we suggest adopting a Modular Closed-Loop Control Architecture for LLM-Driven NPC Behavior, which consists of the generation layer, the fuzzy-symbolic behavioral control stack, the preventive validation layer, and the persistent memory module working in a closed loop. According to our comparison analysis, this design provides a more holistic view of architectural solutions, yet still maintains the flexible nature of the LLM systems. Lastly, we propose three possible research directions for the future: causality-based memory retrieval, reasoning-action integration, and multi-agent coherence. With this survey, we provide an overview of the current state of research, thus giving a unified view on the history of intelligent NPC architectures development. Keywords: Non-Player Characters, Artificial Intelligence, Large Language Models, Game AI, Memory Systems, Retrieval-Augmented Generation, Decision Making, Generative Agents, Behavior Trees, Reinforcement Learning, Finite State Machines.