arXiv cs.CL· Wenbo Pan, Shujie Liu, Xiangyang Zhou, Xianlong Wang, Shiwei Zhang, Bingjing Xu, Wanlu Shi, Xiaohua Jia·· 11 小时前AI 评分38
Mstar：让每个任务拥有自己的记忆框架
M$^\star$: Every Task Deserves Its Own Memory Harness
AI 导读
Mstar 通过反思式代码进化自动为不同任务发现专属记忆程序，将记忆框架实现为定义 Schema、Logic 和 Instruction 的可执行 Python 程序并联合优化。在对话、具身规划和专业推理四类任务上，Mstar 性能均超越现有基线，且进化出的记忆程序在各领域呈现结构上不同的处理机制。该工作入选 NeurIPS 2026 poster。
正文 · 原文
Abstract:Large language model agents rely on a memory harness to write, organize, retrieve, and use past experience. A harness that works for one task often fails on another, because conversations, embodied planning, and expert reasoning require different storage and retrieval behavior. To address this limitation, we introduce Mstar, a method that automatically discovers task-optimized memory programs through reflective code evolution. Specifically, Mstar operationalizes a memory harness as an executable Python memory program that defines the Schema, Logic, and Instruction, and optimizes these components jointly. We evaluate Mstar on four tasks covering conversation, embodied planning, and specialized reasoning. Our results demonstrate that Mstar improves performance over existing baselines across all evaluated tasks. Furthermore, the evolved memory programs exhibit structurally distinct processing mechanisms for each evaluated domain. These findings suggest that every task benefits from its own memory harness, and that memory-program search provides a concrete way to discover it.
| Comments: | NeurIPS 2026 (poster). Code: this https URL ; Live demo: this https URL |
| Subjects: | Programming Languages (cs.PL); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2604.11811 [cs.PL] |
| (or arXiv:2604.11811v3 [cs.PL] for this version) | |
| https://doi.org/10.48550/arXiv.2604.11811 arXiv-issued DOI via DataCite |
Submission history
From: Wenbo Pan [view email]
[v1]
Fri, 10 Apr 2026 03:22:26 UTC (761 KB)
[v2]
Sat, 23 May 2026 08:43:43 UTC (1,318 KB)
[v3]
Fri, 30 Oct 2026 06:13:10 UTC (1,308 KB)
来源：arXiv cs.CL · arxiv.org