Adaptive Budgeted Forgetting Framework for Autonomous Agent Memory Management
arXiv·medium signal
A new framework addresses the memory accumulation problem in long-horizon conversational agents, where uncontrolled growth causes temporal decay and false memory propagation — LOCOMO benchmark shows performance degrading from 0.455 to 0.05 across stages. The adaptive budgeted forgetting approach uses relevance-guided scoring and bounded optimization to regulate what agents retain. MultiWOZ evaluation shows 78.2% accuracy with only 6.8% false memory rate under persistent retention.