Object-Centric Modeling Improves LLM Agent Performance.

As Large Language Model (LLM) agents accumulate experience, managing and reusing free-form textual memories becomes increasingly challenging, often leading to difficulties in validation and maintenance. Existing symbolic approaches learn skills or world models but frequently store local procedures or assume simplified dynamics. This paper introduces Object-Centric Environment Modeling (OCM), an approach that structures an agent's experience into an executable, object-centric environment model. OCM maintains two interconnected codebases: "object knowledge," which defines environment entities and their mechanisms as Python classes, and "procedure knowledge," which records reusable interaction patterns that import and utilize the object model. OCM operates online, reflecting on each trajectory to update both knowledge bases and verify procedure execution against the updated object model. For future interactions, the agent uses progressive knowledge disclosure, inspecting compact code signatures before delving into source code. Experiments demonstrate that OCM achieves the best average rank across benchmarks and significantly reduces invalid actions, highlighting the benefits of building object-centric environment models for agent improvement.