Paper Overview
Research Area: LLM Agents
Authors: Haozhen Zhang, Haodong Yue, Quanyu Long, Jianzhu Bao, Qingyuan Liu, Tao Feng, Bohan Liu, Weida Liang, Wenya Wang
Release Date: 2026-10-05
arXiv: 2610.06830
Summary
Memory has become a core component of the LLM agent ecosystem, enabling information retention and reuse across interactions. However, most existing agent memory systems construct memories in a query-agnostic way, resulting in unnecessary preprocessing costs and discarding details that later prove essential. Recent work has begun shifting memory processing toward runtime adaptation, but typically focuses on specific operations or fixed processing schemes; flexible control over performance, cost, and latency remains largely unexplored.
The authors propose MemPilot, a flexible framework that orchestrates on-demand memory curation under competing objectives. Specifically, a multi-step LLM policy is optimized via reinforcement learning to iteratively choose between retrieving from query-agnostic memory or delegating query-specific curation of raw multimodal history to heterogeneous LLMs and VLMs. The policy jointly controls:
- Evidence quantity
- Curation instructions
- Model selection
- Visual access
- Objective-wise advantage decoupling: advantages are estimated separately for each objective before aggregation, allowing optimization under competing goals.
- Prefix-based marginal utility estimation: provides fine-grained credit assignment across multi-step rollouts.
This enables fine-grained allocation of runtime computation.
Method Highlights
Results
Experiments on five multimodal agent memory benchmarks show that MemPilot achieves favorable performance-cost-latency trade-offs under diverse optimization preferences.
Original Abstract
MemPilot is a flexible framework that orchestrates on-demand memory curation under different performance-cost-latency preferences. An RL-optimized multi-step LLM policy chooses between retrieving query-agnostic memory and delegating query-specific curation to heterogeneous LLMs and VLMs.