dsh-mneme

DeepSeek Harness 的跨会话长期记忆
聊过的项目 · 提过的偏好 · 做过的决定 —— 关掉窗口，它还记着

中文 · English

---

Mneme（Μνήμη）是记忆女神 Mnemosyne 的名字。希腊人把记忆交给她，也把梦交给她——名字落在插件上是同一个意思：白天记住，夜里整理。autoDream 就在你空闲时做这件事，把常看的留在热区，把陈旧的压成摘要、归档。

它解决什么

每次新开对话，AI 都像第一次认识你。你重新讲一遍项目背景，重新交代偏好，重新说服它别改那个已经定下来的决定。

dsh-mneme 给 DeepSeek Harness 一份跨会话的记忆：对话结束，值得留下的部分自动蒸馏进本机；下次开口，相关的几条自己浮上来。记忆不是锁在向量库里的一堆浮点数——它落在本机的一份 SQLite，旁边配一份 Markdown 镜像，你随时能打开、能改，改了不会被机器盖掉。

_{写进来 → 本机存着 → 空闲时整理 → 四路取回 → 送进会话。}

安装

dsh plugin --profile web add @modusensus/dsh-mneme
dsh web

两条硬要求，不满足装不上：

• Node 24+ —— 记忆库用的是 node:sqlite，Node 22 起不来库。
• DSH 0.1.x 或 0.2.x —— 两条线都在兼容范围内；CI 对 0.2 线全绿。

本仓库是单仓多包：可安装的本体在 dsh-mneme/，根目录那份 package.json 只是仓库壳（private、无 dsh 字段）。走 Git 或 tarball 时路径必须指到子目录，否则装到的是壳本身。四种安装来源的写法见完整文档。

它落在你机器的哪几个地方：

目录	是什么
~/.dsh/memory	记忆库本体：SQLite 正本，加十份可手改的 Markdown 镜像
~/.dsh/mneme	本地嵌入模型与运行时缓存，删掉会自动重下
~/.dsh-mneme	CLI / MCP 的连接配置（url 与 token）

想卸掉：

dsh plugin --profile web remove @modusensus/dsh-mneme

插件目录没了，~/.dsh/memory 里的记忆还在——想清干净就自己删这个目录。三个目录名字很像，别认错：目录说明。

你会用到的

• 自动记住 — 对话结束时蒸馏出值得留的事实、偏好、决定，不需要你手动「保存到记忆」。
• 按需注入 — 下次开口，相关的几条自己进上下文；热度、近因、重要度一起算，不是简单按时间堆。
• 空闲巩固 — autoDream 在你不用的时段合并重复、压缩陈旧、归档过期。越用越精炼，不是越堆越大。
• 矛盾不替你做主 — 两条记忆打架时先冻结，并排摆在状态页，等你说保留哪条。
• 记忆是文件 — 十份 Markdown 镜像，能看到、能改、能导出；手改过的内容在合并回库时优先保留。
• 一块面板 — 记录、浏览、筛选、搜索、改、删，都在设置页里。

凭什么信它

• 每条记忆都带着来历 — 从哪段对话来的、什么时候改过、被谁动过。自动整理同样留账：输入快照、决策明细、结果哈希，同一次整理可以复现回放。
• 隔离与离线 — 可按 agent / 工作区隔离；把嵌入换成本地模型后，语义检索完全离线，不需要任何 API Key。
• 删对话 ≠ 删记忆 — 记忆是跨会话的全局知识，清空聊天窗口不影响已经存下的部分。
• 随时能接管 — 记忆库就是本机 SQLite 加一堆 Markdown，没有私有格式把你锁在里面。

对比 Zep / Mem0 / Letta / Hindsight 的 11 行逐项表（存储、检索、写入时抽取、离线巩固、冲突处理、遗忘、作用域、可审计性、多宿主、离线、许可证）：能力对比。「—」不等于「差」。

面板

一块面板装在设置页里：记录、浏览、筛选、搜索、改、删，外加实体关系图谱与状态审计。界面截图单独放在 docs/PANEL.md —— 面板一改，只有那一页需要跟着更新。

接到别处

方式	你能得到什么
MCP	一个零依赖 stdio server（独立包 mneme-memory）；Claude Code / Cursor / Codex / OpenCode 等任何 MCP 客户端都能挂上六个记忆工具
独立 daemon	dsh-mneme-serve 让记忆库常驻：DSH 关着时，脚本、网页端、你自己的面板读写同一份记忆
外部 API	HTTP 接口，别的插件想把记忆写进来，走 API 就行
社区适配器	如 Mneme-Bridge，把 DeepSeek 网页版接进同一个库

隐私承诺

• 只落在本机。 SQLite 存储加一份可手工编辑的 Markdown 镜像，中间没有任何 mneme 服务器。
• 记忆是你能打开的文件。 Markdown，人类可读、可手改，手改过的内容在合并回库时优先保留。
• 不需要 mneme 自己的 API Key。 巩固与摘要走你已配置的宿主模型（autoDream / autoSummarize 默认开）。
• 想连巩固也完全离线：把宿主换成本地模型，并清空 dreamProvider / dreamModel 与 summarizeProvider / summarizeModel（配了这两组会绕过宿主直连，只换宿主不生效），或直接关掉这两个开关。
• 无遥测、无分析、无远程日志。

文档

想找什么	去哪
完整参考：全部配置项、CLI、外部 API、架构	dsh-mneme/README.md
每个配置键的默认值与含义	docs/CONFIGURATION.md
和别的记忆系统的能力对照	docs/COMPARISON.md
每个版本改了什么	CHANGELOG · Releases
三个目录到底谁是谁	docs/DIRECTORIES.md
致谢与相关研究	完整文档 · 致谢

 

"记忆不是把对话存下来，是让下一次对话从上次继续。"

 

_{© 2026 MODUSENSUS · MIT}

 

---

dsh-mneme

Long-term, cross-session memory for DeepSeek Harness
projects · preferences · decisions — still there after you close the window

Mneme (Μνήμη) is the name of Mnemosyne, the goddess of memory — and of dreams. The name means the same thing here: remember by day, tidy up at night. autoDream does that while you are idle: hot memories stay hot, stale ones compress into summaries, old ones archive.

What it is for

Every new chat, the AI meets you for the first time. You explain the project again, restate your preferences, and re-argue a decision that was already settled.

dsh-mneme gives DeepSeek Harness memory that outlives a conversation: when a chat ends, what is worth keeping is distilled into your machine, and the relevant pieces surface on their own next time. The memories are not a pile of floats locked in a vector store — they live in a local SQLite file with a Markdown mirror beside it: open it, read it, edit it, and your edits are not overwritten.

The whole loop in one picture (labels in Chinese): docs/memory-loop.svg.

Install

dsh plugin --profile web add @modusensus/dsh-mneme
dsh web

Two hard requirements — without them it will not install:

• Node 24+ — the store uses node:sqlite; on Node 22 the database will not open.
• DSH 0.1.x or 0.2.x — both lines are in range, and CI is green on the 0.2 line.

This repository is a multi-package repo: the installable plugin lives in dsh-mneme/, and the root package.json is only a repo shell (private, no dsh field). Git and tarball specs must point at the subdirectory, or you install the shell itself. All four install sources are in the full docs (Chinese).

Where it lands on your machine:

Path	What it is
~/.dsh/memory	The store itself: the SQLite original plus ten hand-editable Markdown mirrors
~/.dsh/mneme	Local embedding model and runtime cache; delete it and it refetches
~/.dsh-mneme	CLI / MCP connection settings (url and token)

To remove it:

dsh plugin --profile web remove @modusensus/dsh-mneme

The plugin directory goes away, the memories in ~/.dsh/memory stay — delete that directory yourself if you want them gone. The three names are easy to mix up: directories.

What you get

• Remembered automatically — when a chat ends, the facts, preferences and decisions worth keeping are distilled; there is no "save to memory" step.
• Injected on demand — the relevant handful enters context next time, ranked by heat, recency and importance rather than simply by date.
• Consolidated while idle — autoDream merges duplicates, compresses what went stale, archives what expired. The store gets leaner, not fatter.
• Conflicts are not decided for you — when two memories disagree, both are frozen and shown side by side, and you pick which one stays.
• Memories are files — ten Markdown mirrors you can read, edit and export; hand-edited text wins when it is merged back.
• A panel — record, browse, filter, search, edit, delete, all in the settings page.

Why you can trust it

• Every memory carries its provenance — which conversation it came from, when it changed, who touched it. Consolidation leaves a receipt too: input snapshot, decision detail, result hash, so the same run can be replayed.
• Isolation and offline — memories can be scoped per agent and workspace; point the embedder at a local model and semantic search runs fully offline, with no API key.
• Delete the session ≠ delete the memory — memories are cross-session knowledge; clearing a chat window leaves what was saved alone.
• You can always take it over — the store is a local SQLite file plus Markdown; no proprietary format holds your data.

How does it compare with Zep / Mem0 / Letta / Hindsight? An 11-row table — storage, retrieval, write-time extraction, offline consolidation, conflicts, forgetting, scoping, auditability, multi-host, offline, license: capability comparison (Chinese). "—" does not mean "worse".

The panel

One panel in the settings page: record, browse, filter, search, edit, delete — plus an entity graph and an audit view. Screenshots live in docs/PANEL.md: when the panel changes, only that page needs updating.

Elsewhere

Way in	What you get
MCP	A zero-dependency stdio server (standalone package mneme-memory); Claude Code / Cursor / Codex / OpenCode and any other MCP client can mount the six memory tools
Standalone daemon	dsh-mneme-serve keeps the store running: with DSH closed, scripts, web bridges and your own panels read and write the same memories
External API	An HTTP surface, so another plugin can write memories in through the API
Community adapters	Such as Mneme-Bridge, which brings the DeepSeek web chat into the same store

Privacy

• Local only. A SQLite store plus a hand-editable Markdown mirror; no mneme server sits in between.
• Memories are files you can open. Markdown, human readable and hand editable; your edits are kept when merged back.
• No mneme API key required. Consolidation and summaries go through the host model you already configured (autoDream / autoSummarize, on by default).
• Fully offline consolidation is possible: point the host at a local model, and clear dreamProvider / dreamModel and summarizeProvider / summarizeModel (if set, they bypass the host), or turn the two switches off.
• No telemetry, no analytics, no remote logs.

Docs

Looking for	Where
Full reference: every setting, CLI, external API, architecture	dsh-mneme/README.md
Defaults and meaning of each config key	docs/CONFIGURATION.md
How it compares with other memory systems	docs/COMPARISON.md
What changed in each version	CHANGELOG · Releases
Which of the three directories is which	docs/DIRECTORIES.md
Credits and related work	full docs · credits (Chinese)