LLM-Driven Active Behavioral Forgetting for AI Avatar in Metaverse Environments

Abstract

Behavioral forgetting is a key cognitive function that adaptively suppresses outdated action policies. Despite its importance, its implementation within AI avatars for metaverse environments remains relatively underexplored, particularly in the domain of autonomous agent control primarily due to the difficulty of integrating flexible memory control into existing Reinforcement Learning (RL) frameworks and the practical constraints of real-time deployment. To address this gap, we introduce Forgetting-Enabled Avatar Control (FEAC), a dual-model architecture that integrates a large language model (LLM)-based memory evaluator with a policy network based on Proximal Policy Optimization (PPO). In FEAC, forgetting decisions are made via in-context learning and implemented through an adaptive observation gating mechanism. We evaluate FEAC in a virtual warehouse exploration task across nine forgetting regimes over 5 × 106 steps. The results reveal a U-shaped trade-off where moderate forgetting achieves an optimal balance between exploration efficiency and behavioral stability. Furthermore, comparative analysis demonstrates that FEAC outperforms a heuristic rule-based baseline by 17% in optimal regimes and maintains robust performance even under aggressive suppression schedules. Additionally, FEAC reduces redundant revisits, stabilizes reaction times, and maintains compact, goal-directed trajectories. These outcomes demonstrate that neuroscience-inspired forgetting mechanisms can enhance avatar efficiency and adaptability, charting a path toward broader metaverse applications such as virtual training and navigation.