学习如何记忆：结构化和可迁移代理记忆的元认知管理方法
人工智能
2026-01-13 v1
摘要
大型语言模型（LLM）代理越来越依赖累积的记忆来解决长期决策任务。然而，大多数现有方法将记忆存储在固定表示中，并在单个或隐式的抽象层次上重用，这限制了泛化能力，常常导致在分布迁移时出现负迁移。本文提出了元认知记忆抽象方法（MCMA），将记忆抽象视为可学习的认知技能，而非固定的设计选择。MCMA通过结合冻结的任务模型和已学习的记忆副手，将任务执行与记忆管理解耦。记忆副手使用直接偏好优化进行训练，它决定记忆应如何结构化、抽象和重用。记忆进一步组织为抽象层次的层次结构， enables selective reuse based on task similarity. When no memory is transferable, MCMA transfers the ability to abstract and manage memory by transferring the memory copilot. Experiments on ALFWorld, ScienceWorld, and BabyAI demonstrate substantial improvements in performance, out-of-distribution generalization, and cross-task transfer over several baselines.
引用
@article{arxiv.2601.07470,
title = {Learning How to Remember: A Meta-Cognitive Management Method for Structured and Transferable Agent Memory},
author = {Sirui Liang and Pengfei Cao and Jian Zhao and Wenhao Teng and Xiangwen Liao and Jun Zhao and Kang Liu},
journal= {arXiv preprint arXiv:2601.07470},
year = {2026}
}