MemArbiter: Decision-Time Memory Arbitration for Long-Horizon LLM Agents
2026-08-03 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors found that large language model agents sometimes struggle to use important past information to make good decisions in long tasks, even when they can access this information. They call this problem the Memory-Action Gap. To fix it, they created MemArbiter, a system that breaks down memory into smaller parts and organizes it by function so the agent can better choose what information to focus on when acting. Their tests show MemArbiter helps agents succeed more often and avoid repeating mistakes. Overall, the authors show that smart organization of memory leads to better decision-making.
Large Language ModelsMemory ManagementLong-Horizon TasksMemory-Action GapMemory ArbitrationAtomic ItemsMemory BanksAction-Generation ModelsToken BudgetsALFWorld
Authors
Jiajun Dong, Yutao Hu, Fengrui Fan, Shihan Dou, Yueming Wu, Deqing Zou
Abstract
Large language model (LLM) agents must retain and use cross-step information to act coherently in long-horizon tasks. Existing methods improve memory accessibility, yet action-relevant information may still fail to guide the current decision because it is poorly formed, organized, prioritized, or presented. We call this post-access failure the Memory-Action Gap. We propose MemArbiter, a function-aware memory arbitration framework that addresses the memory-management-induced component of this gap. MemArbiter decomposes interaction histories into atomic items, organizes them into five functional Memory Banks, and combines bank-level demand, item-level relevance, focal-ambient representations, and a temporal presentation gate to dynamically control memory salience. We evaluate MemArbiter on ALFWorld against Flat Retrieval and Flat Recency under unified per-step memory budgets. With an open-weight action-generation model, MemArbiter achieves success rates of 82.8% and 92.5% under 500- and 750-token budgets, outperforming the strongest baseline by 20.9 and 25.4 percentage points, respectively. It also improves post-failure recovery and reduces failed-action repetition and state-action recurrence. These results show that function-aware memory arbitration enables accessible information to guide actions more effectively.