Metacognition is all you need? Using Introspection in Generative Agents to Improve Goal-directed Behavior

Recent advances in Large Language Models (LLMs) have shown impressivecapabilities in various applications, yet LLMs face challenges such as limitedcontext windows and difficulties in generalization. In this paper, we introducea metacognition module for generative agents, enabling them to observe theirown thought processes and actions. This metacognitive approach, designed toemulate System 1 and System 2 cognitive processes, allows agents tosignificantly enhance their performance by modifying their strategy. We testedthe metacognition module on a variety of scenarios, including a situation wheregenerative agents must survive a zombie apocalypse, and observe that our systemoutperform others, while agents adapt and improve their strategies to completetasks over time.