Search papers, labs, and topics across Lattice.
This paper introduces MemArbiter, a function-aware memory arbitration framework designed to enhance decision-making in long-horizon tasks for large language model (LLM) agents by addressing the Memory-Action Gap. By decomposing interaction histories into atomic items and organizing them into five functional Memory Banks, MemArbiter dynamically controls memory salience through a combination of demand, relevance, and temporal presentation. Evaluated on the ALFWorld benchmark, MemArbiter significantly outperforms existing methods, achieving success rates of 82.8% and 92.5% under constrained memory budgets, while also improving recovery from failures and reducing action repetition.
MemArbiter closes the Memory-Action Gap, boosting LLM decision-making success rates by over 20% in long-horizon tasks.
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.
