---
id: 20261003-T0-05
title: "长周期语言代理的记忆使用呈重尾分布，仅看任务成功率会遗漏关键问题"
title_en: "Long-Horizon Language Agents Show Heavy-Tailed Memory Traces, Task Success Alone Misses Key Issues"
url: https://ai.daily.yangsir.net/daily/20261003-T0-05
issue_date: 2026-10-03
publish_date: 2026-10-02T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2610.00010
---

# 长周期语言代理的记忆使用呈重尾分布，仅看任务成功率会遗漏关键问题

一篇新论文指出，长周期语言代理越来越依赖外部记忆作为冻结的世界模型，但当前记忆系统通常只用任务成功率或token成本来评估。作者认为，缺失的评估维度是记忆使用的形态。在有限上下文条件下，记忆访问模式呈现重尾分布。这意味着仅关注任务是否成功，会忽略代理在记忆使用上的结构性特征。该研究为评估和改进语言代理的记忆系统提供了新的分析视角。

## English Version

**Long-Horizon Language Agents Show Heavy-Tailed Memory Traces, Task Success Alone Misses Key Issues**

A new paper argues that long-horizon language agents increasingly rely on external memory as a frozen world model, yet current memory systems are typically judged only by task success or token cost. The authors contend that the missing evaluation dimension is the shape of memory use. Under finite context, memory access patterns exhibit heavy-tailed distributions. This means focusing solely on task success overlooks structural characteristics of how agents use memory. The research offers a new analytical lens for evaluating and improving memory systems in language agents.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2610.00010)

**详情页**：https://ai.daily.yangsir.net/daily/20261003-T0-05

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*