dsh-project-memory

English | 中文

---

English

A DeepSeek Harness plugin that lets the AI learn from each project — recording pitfalls and user coding habits, then injecting them into future conversations so the AI gets smarter the more you use it.

What it does

When you hit a common error, show a coding habit, or explicitly ask the AI to remember something, the AI calls the memory_write tool to save it. Every future conversation in that project (or across all projects, for user-scoped memories) automatically receives those memories as context, so the AI builds on past lessons without being told again.

Three contributions:

Contribution	Who uses it	What it does
memory_write tool	the AI model	Records a pitfall, habit, preference, or fact
agent/pre-step injection	automatic	Prepends all memories to every model request
/memory command	you (human)	View, add, edit, delete, and clear memories

Two scopes

• User memory (user) — applies to all projects. Lives in <DSH_HOME>/memory.yaml (~/.dsh/memory.yaml by default).
• Project memory (project) — applies to the current project only. Lives in <projectRoot>/.dsh/memory.yaml, where project root is the nearest .git ancestor directory.

Both scopes are human-editable (the YAML files are plain text) and can be managed via the /memory command.

Categories

Every memory has one of four categories:

Category	Meaning	Example
pitfall	A common error and its fix	"on Windows, use file:// URLs for ESM imports"
habit	A user coding convention	"all code comments in Chinese"
preference	A workflow choice	"run tests before pushing"
fact	Important project context	"this project requires Node >= 22"

The /memory command

/memory                                     list all memories (user then project)
/memory list user                           list only user-scoped memories
/memory list project                        list only project-scoped memories
/memory add user habit: all comments in CN  add a user-scoped memory
/memory add project pitfall: never push main add a project-scoped memory
/memory edit 2 new content here             edit the memory at index #2
/memory delete 1                            delete by index (from /memory list)
/memory delete <id>                         delete by full UUID (searches both scopes)
/memory clear project                       clear all project memories
/memory clear                               clear all memories (both scopes)

Indices are continuous: user memories come first (#1..), then project memories. The same index works for edit and delete.

Install

From npm (recommended)

npx @deepseek-ai/dsh plugin --profile web add @xiaozhugez/dsh-project-memory

Prebuilt code — installs in one step, no build permission needed.

From GitHub

npx @deepseek-ai/dsh plugin --profile my-profile add github:xiaozhugez/dsh-project-memory

A git install fetches source and runs the prepare script (tsdown) to build lib/index.js. pnpm ≥10 refuses to run a git dependency's build script until it is explicitly allowed, so the first add fails. Copy the package key pnpm prints into that profile's pnpm-workspace.yaml:

allowBuilds:
  dsh-project-memory: true

Then re-run the add command. Treat this as permission to execute the package's code on your machine at install time — only allow packages whose source you trust, and pin a commit (github:xiaozhugez/dsh-project-memory#<sha>) so a later push cannot silently change what runs.

From local directory

npx @deepseek-ai/dsh plugin --profile my-profile add ./project-memory-plugin

Uninstall

npx @deepseek-ai/dsh plugin --profile web remove @xiaozhugez/dsh-project-memory

Verify

npx @deepseek-ai/dsh --profile web --dump-config   # confirm project-memory is in the tree
npx @deepseek-ai/dsh web                             # start using it

Config

Override via a profile or home patch above this bundle:

- id: project-memory
  name: '@xiaozhugez/dsh-project-memory'
  config:
    memoryDir: .dsh      # directory under project root; default .dsh
    autoInject: true      # inject memories into model context; default true

Manual editing

Both memory files are plain YAML and safe to edit directly:

• User: ~/.dsh/memory.yaml
• Project: <projectRoot>/.dsh/memory.yaml

memories:
  - id: some-uuid
    category: pitfall
    content: never commit directly to main
    createdAt: '2026-08-19T10:00:00.000Z'

The in-memory cache checks the file mtime on every read, so manual edits are picked up on the next request within the same session — no restart needed.

Development

This plugin is developed outside the packages/ workspace as an independent plugin, following the Harness plugin development guide.

Build the runtime artifact from source:

pnpm install   # runs prepare → tsdown → lib/index.js
pnpm build     # re-run the build alone

To test during development, run from the harness repo root with the dev overlay (which mounts the TypeScript source directly through tsx):

pnpm dsh web --patch ./project-memory-plugin/cordis.yml

Then open http://127.0.0.1:3080. The plugin's memory_write tool and /memory command are available in the session.

Files

project-memory-plugin/
├── package.json           # dsh bundle manifest (dsh.bundle.patch) + deps
├── pnpm-workspace.yaml     # autoInstallPeers:false (dsh peers come from the host)
├── tsdown.config.ts        # standalone build: src/ → lib/index.js (no type-check)
├── cordis.patch.yml         # installable bundle patch (uses package name)
├── cordis.yml              # dev overlay (uses absolute source path)
├── src/
│   ├── index.ts            # three registrations: tool + command + pre-step
│   └── store.ts            # dual-scope storage, mtime cache, YAML persistence
└── tests/
    ├── command.spec.ts      # store layer: loadAllMemories/updateMemory/clearMemories/mtime
    └── dispatch.spec.ts     # command dispatch: list/add/edit/delete/clear

---

中文

一个 DeepSeek Harness 插件, 让 AI 越用越懂你 —— 记录你踩过的坑和代码习惯,在以后的对话中自动注入, 让 AI 在同一个项目里越用越聪明。

功能介绍

当你遇到一个常见报错、展现出某种代码习惯,或明确要求 AI 记住某件事时,AI 会调用 memory_write 工具把它记下来。之后在该项目(或所有项目,取决于记忆范围)的每次对话中,这些记忆会自动作为上下文注入,AI 就能在过往经验上继续积累,而不必你反复说明。

三个组成部分:

组成	使用者	作用
memory_write 工具	AI 模型	记录一个踩坑、习惯、偏好或事实
agent/pre-step 注入	自动触发	把所有记忆追加到每次模型请求
/memory 命令	你(人类)	查看、添加、编辑、删除、清空记忆

两种范围

• 用户级记忆(user)—— 适用于所有项目。存储在 <DSH_HOME>/memory.yaml(默认 ~/.dsh/memory.yaml)。
• 项目级记忆(project)—— 仅适用于当前项目。存储在 <项目根>/.dsh/memory.yaml,项目根为最近一层含 .git 的目录。

两种范围都可以直接手动编辑(YAML 纯文本),也可以通过 /memory 命令管理。

记忆分类

每条记忆属于以下四类之一:

分类	含义	示例
pitfall	常见错误及其解决办法	"Windows 上 ESM 导入要用 file:// URL"
habit	用户的代码习惯	"所有代码注释用中文"
preference	工作流偏好	"推送前先跑测试"
fact	重要的项目上下文	"本项目要求 Node >= 22"

/memory 命令

/memory                                     列出所有记忆(用户级在前,项目级在后)
/memory list user                           只列出用户级记忆
/memory list project                        只列出项目级记忆
/memory add user habit: 所有注释用中文        添加用户级记忆
/memory add project pitfall: 不要直接推 main  添加项目级记忆
/memory edit 2 新的内容                       编辑编号 #2 的记忆
/memory delete 1                            按编号删除(/memory list 显示的序号)
/memory delete <id>                         按 UUID 删除(两个范围都搜)
/memory clear project                       清空项目级记忆
/memory clear                               清空所有记忆(两个范围)

编号是连续的:用户级记忆排前面(#1..),项目级排后面。edit 和 delete 用的是同一套编号。

安装

从 npm 安装(推荐)

npx @deepseek-ai/dsh plugin --profile web add @xiaozhugez/dsh-project-memory

已预构建,一步安装,无需构建授权。

从 GitHub 安装

npx @deepseek-ai/dsh plugin --profile my-profile add github:xiaozhugez/dsh-project-memory

Git 安装会拉取源码并运行 prepare 脚本(tsdown)构建出 lib/index.js。 pnpm ≥10 要求显式授权才会运行 git 依赖的构建脚本,所以第一次 add 会失败。 把 pnpm 打印的包名填进该 profile 的 pnpm-workspace.yaml:

allowBuilds:
  dsh-project-memory: true

然后重新运行 add 命令。请注意:这等于授权该包在你的机器上、安装时执行 它的代码。只对你信任其源码的包授权,并固定 commit (github:xiaozhugez/dsh-project-memory#<sha>),防止后续推送悄悄改变执行内容。

从本地目录安装

npx @deepseek-ai/dsh plugin --profile my-profile add ./project-memory-plugin

卸载

npx @deepseek-ai/dsh plugin --profile web remove @xiaozhugez/dsh-project-memory

验证

npx @deepseek-ai/dsh --profile web --dump-config   # 确认 project-memory 出现在配置树中
npx @deepseek-ai/dsh web                             # 启动使用

配置

通过 profile 或 home patch 在本 bundle 之上覆盖:

- id: project-memory
  name: '@xiaozhugez/dsh-project-memory'
  config:
    memoryDir: .dsh      # 项目根下的存储目录;默认 .dsh
    autoInject: true      # 是否注入记忆到模型上下文;默认 true

手动编辑

两个记忆文件都是纯 YAML,可以直接编辑:

• 用户级:~/.dsh/memory.yaml
• 项目级:<项目根>/.dsh/memory.yaml

memories:
  - id: some-uuid
    category: pitfall
    content: never commit directly to main
    createdAt: '2026-08-19T10:00:00.000Z'

内存缓存在每次读取时检查文件 mtime,所以手动编辑后同一会话内的下一次请求 就能生效 —— 无需重启。

开发

本插件在 packages/ 工作区之外独立开发,遵循 Harness 插件开发指南。

从源码构建运行时产物:

pnpm install   # 会自动运行 prepare → tsdown → lib/index.js
pnpm build     # 单独重新构建

开发时测试,在 harness 仓库根目录用 dev overlay 运行(通过 tsx 直接加载 TypeScript 源码):

pnpm dsh web --patch ./project-memory-plugin/cordis.yml

然后打开 http://127.0.0.1:3080。会话中即可使用 memory_write 工具和 /memory 命令。

文件结构

project-memory-plugin/
├── package.json           # dsh bundle 清单 (dsh.bundle.patch) + 依赖
├── pnpm-workspace.yaml     # autoInstallPeers:false(dsh peer 由宿主提供)
├── tsdown.config.ts        # 独立构建:src/ → lib/index.js(不做类型检查)
├── cordis.patch.yml         # 可安装的 bundle patch(用包名引用)
├── cordis.yml              # 开发 overlay(用源码绝对路径)
├── src/
│   ├── index.ts            # 三个注册点:工具 + 命令 + pre-step 注入
│   └── store.ts            # 双范围存储、mtime 缓存、YAML 持久化
└── tests/
    ├── command.spec.ts      # 存储层:loadAllMemories/updateMemory/clearMemories/mtime
    └── dispatch.spec.ts     # 命令分发:list/add/edit/delete/clear