Four practical limits of dsh-auto-memory 1.3.0
In a three-week field test of dsh-auto-memory 1.3.0, the author found that automatic summarization is off by default, injected memories are truncated at a byte-budget limit, and unverified candidate records can still be injected. The review also reports weak performance on long-document retrieval. The author considers the tool best suited to manually maintained, single-machine, project-level memory.
The review identifies configuration, truncation and verification requirements that can determine whether the tool fits a user's workflow. It also distinguishes the tool's stated strength in project-level memory from its reported weakness in long-document retrieval and cross-machine sharing.
我踩过的四个记忆坑，从预算到换机
我给自己的 agent 装持久记忆，动机很朴素：同一个项目反复讲背景太费劲。装上 dsh-auto-memory 之后，前三周我踩了四个坑，每一个都不是「装不上」，而是「装了但没按我以为的方式工作」。下面按我踩到的顺序写，不讲教科书分类。
先说它现在长什么样
我实测的版本是 1.3.0（npm 2026-10-29 发布），仓库 122 星（GitHub API，2026-11-05 现查），站内给的是已验证 · L5 运行时冒烟。装完的记忆存在 $DSH_HOME/memory/：一条记忆一个 Markdown 文件，加一个 MEMORY.md 索引，写进系统提示词的是索引而不是全文。
我最开始以为这是「装完就自动记住」的东西。不是。
坑一：以为它会自己记
第一周我聊了三个晚上，第二天开新会话问它「我上次说过什么」，它一脸茫然。翻 README 才看到 autoSummarize 默认 false——不开这个开关，它只在两种情况下写盘：我明确说「记住」，或者模型主动调 memory_write。