This paper provides a comprehensive overview of large language model (LLM)-based game agents (LLMGAs) from a holistic perspective. It introduces the conceptual architecture of LLMGAs, focusing on six essential functional components: perception, memory, thinking, role-playing, action, and learning. The paper surveys existing LLM-based game agents across six genres: adventure, communication, competition, cooperation, simulation, and crafting & exploration. It also presents an outlook on future research directions in this field. A curated list of relevant papers is available at https://github.com/git-disl/awesome-LLM-game-agent-papers. The paper discusses the architecture of LLMGAs, including perception, memory, role-playing, thinking, action, and learning. Perception involves capturing game state information from various modalities. Memory stores past experiences and knowledge for future use. Role-playing enables agents to simulate specific roles within the game. Thinking involves reasoning and planning for informed decision-making. Action translates generated decisions into executable actions. Learning continuously improves the agent's abilities through experience. The paper categorizes existing studies into six game genres and highlights key findings and methodologies in each. For adventure games, text-based and video-based games are discussed, with examples like Zork I and Red Dead Redemption 2. Communication games involve negotiation and deception, with examples like Werewolf and Diplomacy. Competition games test skill and strategy, with examples like StarCraft II and Pokémon battles. Cooperation games emphasize teamwork, with examples like Overcooked and Minecraft. Simulation games replicate real-world events, with examples like The Sims and Civilization. Crafting & exploration games involve gathering resources and building structures, with examples like Minecraft and Crafter. The paper also discusses various approaches to LLMGAs, including in-context feedback learning, supervised fine-tuning, and reinforcement learning. It highlights the importance of grounding LLMs in environments to enable learning and development. The paper concludes that LLMGAs have the potential to advance towards Artificial General Intelligence (AGI) by enabling human-like decision-making and cognitive abilities. The paper emphasizes the need for further research and development in this emerging field.This paper provides a comprehensive overview of large language model (LLM)-based game agents (LLMGAs) from a holistic perspective. It introduces the conceptual architecture of LLMGAs, focusing on six essential functional components: perception, memory, thinking, role-playing, action, and learning. The paper surveys existing LLM-based game agents across six genres: adventure, communication, competition, cooperation, simulation, and crafting & exploration. It also presents an outlook on future research directions in this field. A curated list of relevant papers is available at https://github.com/git-disl/awesome-LLM-game-agent-papers. The paper discusses the architecture of LLMGAs, including perception, memory, role-playing, thinking, action, and learning. Perception involves capturing game state information from various modalities. Memory stores past experiences and knowledge for future use. Role-playing enables agents to simulate specific roles within the game. Thinking involves reasoning and planning for informed decision-making. Action translates generated decisions into executable actions. Learning continuously improves the agent's abilities through experience. The paper categorizes existing studies into six game genres and highlights key findings and methodologies in each. For adventure games, text-based and video-based games are discussed, with examples like Zork I and Red Dead Redemption 2. Communication games involve negotiation and deception, with examples like Werewolf and Diplomacy. Competition games test skill and strategy, with examples like StarCraft II and Pokémon battles. Cooperation games emphasize teamwork, with examples like Overcooked and Minecraft. Simulation games replicate real-world events, with examples like The Sims and Civilization. Crafting & exploration games involve gathering resources and building structures, with examples like Minecraft and Crafter. The paper also discusses various approaches to LLMGAs, including in-context feedback learning, supervised fine-tuning, and reinforcement learning. It highlights the importance of grounding LLMs in environments to enable learning and development. The paper concludes that LLMGAs have the potential to advance towards Artificial General Intelligence (AGI) by enabling human-like decision-making and cognitive abilities. The paper emphasizes the need for further research and development in this emerging field.