MIT 协议的 DSH 原生插件，为 DeepSeek Harness 的 AI 提供跨会话 agent memory——`dsh plugin add` 一键安装，自带 cordis.patch.yml 自动注册，无需手动配置。
DSH 适配
原生运行时
作者声明
安全审计
未审计
最后核验
2026-08-21
许可证
MIT
01它能帮你完成什么？
Give DeepSeek Harness agents cross-session long-term agent memory that survives new sessions and days apart
Five memory tracks — user profile, global facts, project key memories (auto-injected, optionally scoped to a single git branch), project logs and daily logs — plus emotional-feedback records you can ask the agent to analyze later
Developers running DeepSeek Harness (dsh) who are tired of re-explaining project context every time they open a new session
Share project agent memory across multiple computers so a switch of machines never loses context
Project key memories, project logs, archives and project todos consistent across devices; with a shared memory repo, global memory (user profile / daily logs / todos) syncs too — stored in dedicated git branches that never touch your code
People working on the same project from two or more computers, or teammates collaborating on one project
Let the agent's skills and memory library evolve themselves so repeated lessons become reusable process
Skills auto-solidified from recurring mistakes and executed as fixed workflows next time; optional turn-level memory review where the agent proposes memories for your confirmation every N turns
Heavy dsh users who want their agent memory and skills to compound over time without hand-maintaining them
02如何接入 DeepSeek Harness？
前置条件
- A working DeepSeek Harness (dsh) installation with at least one profile (the README walks through the `web` profile as the example)
安装步骤
- 01
Run `dsh plugin --profile web add github:csyangwen/dsh-memory-evolve` to install into the profile (local `link:` paths and git/registry package addresses also work)
$ dsh plugin --profile web add github:csyangwen/dsh-memory-evolve
- 02
Restart dsh web to activate — the plugin bundles its own cordis.patch.yml (dsh.bundle.patch declaration), so the host auto-registers it with no manual config
验证接入成功
- Open any session after install: a new row of capability tabs appears (记忆 · 技能 · 待办 · 无限画板 · COI 调度 · 会话广播 · 提示词 · 记忆同步 · 模型设置 · 书签 · 会话评审 · Web UI 设置 · Memory Evolve 设置)
回滚
- Temporarily disable without uninstalling: add `- id: dsh-memory-evolve` with `disabled: true` to the profile's cordis.patch.yml (useful when the plugin breaks DSH startup)
- Fully uninstall with `dsh plugin --profile web remove dsh-memory-evolve` — all effects are cleaned up automatically with the plugin
03DSH 适配与能力边界
Native DSH plugin installed via `dsh plugin add`; ships its own cordis.patch.yml (dsh.bundle.patch declaration) so the host auto-registers it with no manual config
Five-track cross-session agent memory
normal conversations with the dsh agent, plus explicit "记一下…" / "查一下记忆…" requests→user profile, global facts, project key memories (auto-injected, git-branch-scoped if needed), project logs and daily logs — progress auto-recorded every turn
writes memory and log entries to disk; project logs are tagged with the current git branchagent-proposed memories/todos/skills only take effect after your confirmationGit-based memory sync across devices
per-project sync toggle in the 记忆同步 Tab; optional shared memory repo address→memories stored in dedicated git branches (never polluting code) and shared across computers, with conflict resolution (keep local / keep remote / keep both)
pushes memory branches to your own code repo or a designated shared memory repo when you click 同步并推送Skill self-evolution & skill manager
recurring mistakes/methods discovered during work; custom skill directories→methods auto-solidified into skills and reused on similar tasks; browse / search / disable / edit skills from the 记忆 Tab
new skills only take effect after your confirmation; the former standalone dsh-skills-manager plugin must be uninstalled before upgradingTurn-level memory review (opt-in)
reviewEnabled / reviewInterval overrides in the profile's cordis.patch.yml→every N user turns the agent distills memory-worthy facts and hands them to you for confirmation, so the memory library grows on its own
disabled by default; consumes extra model turns while reviewing
04适合谁？何时不该用？
适合
- Developers running DeepSeek Harness (dsh) who are tired of re-explaining project context every time they open a new session
- People working on the same project from two or more computers, or teammates collaborating on one project
- Heavy dsh users who want their agent memory and skills to compound over time without hand-maintaining them
不适合
- Do not manually insert the plugin into the profile's cordis.patch.yml: the bundled patch already auto-registers it, and a duplicate insert with the same id makes the loader fail with 'duplicate loader entry id' so DSH cannot start. Upgraders from old docs must delete the legacy insert line.
05兼容性、维护与安全提示
- Do not manually insert the plugin into the profile's cordis.patch.yml: the bundled patch already auto-registers it, and a duplicate insert with the same id makes the loader fail with 'duplicate loader entry id' so DSH cannot start. Upgraders from old docs must delete the legacy insert line.
- Many capabilities (memory review, session advisor, canvas, notify, sync module) are disabled by default to keep the agent's tool list clean; you must enable them in the 'Memory Evolve 设置' panel before they work.
- Memory sync is opt-in, but once enabled memories are pushed to a git remote — by default a dedicated branch in your own code repo, or a shared memory repo if you configure one. Only projects with sync turned on are affected; everything else stays purely local.
MIT · actively maintained (last push 2026-08-17)
06常见问题
dsh-memory-evolve 如何接入 DeepSeek Harness？是原生插件还是 MCP？
它是 DSH 原生插件。运行 `dsh plugin --profile web add github:csyangwen/dsh-memory-evolve`（本地 link:、git 或 registry 包地址也可以），然后重启 dsh web。插件包自带 cordis.patch.yml（dsh.bundle.patch 声明），host 端自动注册，不走 MCP，也无需手动配置。
安装后怎么确认插件生效了？
打开任意会话，应看到新出现的一排能力标签：记忆 · 技能 · 待办 · 无限画板 · COI 调度 · 会话广播 · 提示词 · 记忆同步 · 模型设置 · 书签 · 会话评审 · Web UI 设置 · Memory Evolve 设置。
AI 会擅自往记忆里写东西吗？
不会。AI 自建的记忆、待办、技能都先进待确认队列，真实改变 AI 行为的写入必须你确认后才生效；只有每回合的项目日志/每日日志进展是自动记录的。
记忆数据会传到网络上吗？
默认只存本地。开启「记忆同步」后，记忆会推送到你代码仓库的专属分支（或你配置的共享记忆仓库），记忆与代码彻底隔离；没开同步的项目保持纯本地状态。
为什么不能手动往 cordis.patch.yml 里 insert 本插件？
bundle patch 已自动注册，手动 insert 同 id 会让加载器报 duplicate loader entry id，DSH 无法启动。改配置（如开启 reviewEnabled）请用 id 顶层覆盖而非 insert；想临时禁用，给该 id 加一行 `disabled: true` 即可，无需卸载。
07相关的 DSH 工作流
weknora
作者 tencent
开源的 LLM 知识库平台：把原始文档转化为可检索的 RAG 知识库、自主推理智能体和自动维护的 Wiki 系统，支持多租户部署。
honcho
作者 plastic-labs
用于构建有状态 AI 智能体的记忆库，支持跨会话连续学习与上下文工程。
mirage
作者 strukto-ai
全球首个面向 AI 智能体的统一虚拟文件系统。
reme
作者 agentscope-ai
ReMe：面向 AI Agent 的记忆管理套件——帮 Agent「记住我、优化我」，支持 RAG 与长期记忆。
08数据与来源
插件包内自带 `cordis.patch.yml`（`dsh.bundle.patch` 声明），`dsh plugin add` 安装后 **host 端自动注册，无需任何手动配置**。以 web profile 为例，两步装好：
# 1. 安装到 profile（本地目录用 link:，也可用 git/registry 包地址） dsh plugin --profile web add github:csyangwen/dsh-memory-evolve # 2.…
页面基于项目公开文档、仓库元数据和 DSH Plugins 的结构化解析生成；最后核验于 2026-08-21。发现错误？提交更正。
最佳 DeepSeek Harness 插件
从全目录挑出的 12 个值得优先安装的插件，覆盖各个分类。