dsh-long-memory

给 DeepSeek Harness 的自动长期记忆插件：每轮对话结束后自动提炼值得跨会话记住的事实，新会话自动带回来。

English · 更新日志

---

中文

特性

• 自动提炼：每轮对话正常结束后，把这一轮的用户消息与助手最终回复交给当前会话使用的模型，要求输出不超过 3 条稳定事实。不需要手动说「记住这个」。
• 全局 / 项目两级：global 记用户身份、偏好、禁忌、语言与风格；workspace 记当前目录的技术栈、命令、约定、关键决定与踩过的坑，互不污染。
• 自动去重合并：与已有条目高度重合时合并更新时间并保留更完整的表述，而不是无限堆积；单工作区上限 500 条，超出淘汰最久未更新的。
• 自动带回：新会话首次组装系统提示时，注入「全局 + 当前工作区」的最近记忆摘要（≤ 1400 字）。同一会话内摘要冻结，保证系统提示不变，避免每轮打穿 prompt 缓存。
• 三个工具：memory_search、memory_save、memory_forget，覆盖自动提炼够不到的手动查询与纠错。
• 纯本地、零依赖：数据是一个普通 JSON 文件，不联网、不引入任何第三方 npm 包。

环境要求

• DeepSeek Harness 0.2.x（已在 0.2.1-rc.3 的 desktop profile 上验证）。
• 插件只跑在 Host 侧，无 Client 代码，不需要构建。

安装

1. 把本仓库下载到本地任意目录（git clone 或下载 ZIP 解压）。
2. 在 DSH 里打开 设置 → 插件，选择从本地目录安装，并指向该目录。 等价的自动化方式：调用 plugin_manager 的 install_bundle，target 传该目录的绝对路径。
3. 安装返回 application: applied 后，在插件列表里能看到「长期记忆 / Long-term Memory」条目。

插件没有配置项，装上即用。安装后无需重启；status.json 等诊断文件从 0.1.1 起才有。

使用

自动进行，你什么都不用做：正常对话，下一轮（或下一个会话）它就知道该记什么、该带回什么。

三个工具只在需要时调用：

工具	作用	参数
memory_search	按关键词检索记忆，返回条目 id	query（必填）、scope = auto/global/workspace、limit
memory_save	手动存一条稳定事实，近似重复会合并	text（必填）、scope = global/workspace、tags
memory_forget	按 id 删除一条记忆	id（必填）

详情面板：点左侧栏的长期记忆图标，打开整页面板（设置里也保留同一页）。面板读取实时记忆库并给出分析：总条数 / 全局 / 工作区 / 近 7 天新增，健康度评分（重复条目、超过 30 天未更新、上次提炼失败都会扣分并列出原因），Top 标签，上次提炼的结果与耗时，以及一份支持搜索、按范围过滤、逐条删除的完整列表。

每条记忆都能展开 「查看对话」：显示产生它的那一轮完整用户消息与助手回复（含会话 id、轮次、时间），以及这条记忆的创建/更新时间、命中次数、来源与工作区完整路径。来源对话在提炼时随记忆一起保存，每条最多保留最近 5 个来源。

日常日记：面板顶部有「记忆 / 日记」两个标签。日记每天一篇，自动从当天的记忆里写成第一人称短文，底部附一条「大肥鱼」的情绪留言，并记录当天观察到的用户日常习惯（习惯会作为 global 记忆、带 habit 标签继续参与后续提炼与检索）。日记存在 diary.json；想看最新的点「重写今天」。

记忆范围怎么选：

• 用户本人的习惯、称呼、语言、禁忌、审美偏好 → global，所有项目通用。
• 只对当前目录成立的：语言版本、包管理器、构建命令、目录约定、架构决定 → workspace。

数据在哪：

%DSH_HOME%\plugin-data\dsh-long-memory\      # Windows 默认 C:\Users\<你>\.dsh\...
~/.dsh/plugin-data/dsh-long-memory/          # macOS / Linux
├── memories.json    # 记忆库（可以直接用编辑器看，格式是普通 JSON）
├── last-run.json    # 最近一次提炼的明细：新增/合并条数、耗时、错误
└── status.json      # 本次启动以来的计数：看到的轮数、成功/失败次数、最近错误

DSH_HOME 环境变量可以覆盖根目录。删除 memories.json 等于清空全部记忆。

成本：每完成一轮会多一次小请求（输入约 2k tokens，输出上限 700）；系统提示里多一段 ≤1400 字的摘要，且每个会话只计算一次，对缓存友好。

隐私：记忆是从对话中提炼的，提炼请求发给当前会话已经在用的模型路由；记忆文件只存在本地，插件本身不上传任何东西。

工作原理

1. 监听 session/event：turn/start 开缓冲区，收集 user/message 与 assistant/message 的文本。
2. turn/end 且 reason = completed 时，用 session.requestHeader() 拿到当前路由，请求模型输出 JSON： [{"scope":"global|workspace","text":"…","tags":["…"]}]。 子代理会话（header.origin = subagent）和非正常结束的轮次直接跳过。 如果安装/重载发生在轮次进行中（错过了 turn/start），会从会话日志回放该轮内容。
3. 解析、清洗（单条 ≤ 300 字、最多 3 条），与已有条目做包含式相似度比较：≥ 0.82 视为同一条，合并且保留更长的表述；否则新增。
4. 新会话第一次组装系统提示时，把全局与当前工作区的记忆按更新时间排序、拼成摘要注入；同一会话内不再变化。
5. 中文相似度用「字符 bigram + 英文单词」的集合计算，Jaccard 与包含度取较大值，所以「用户喜欢深色模式」和「用户偏好深色模式」能被判为同一条。

卸载 / 关闭

在插件管理里禁用或移除 dsh-long-memory 即可，记忆文件会保留；重新启用后继续使用。彻底清除请删除 %DSH_HOME%\plugin-data\dsh-long-memory\。

常见问题

• 新会话没看到记忆摘要：摘要按会话冻结，先确认 memories.json 里确实有内容；空库不会注入任何东西。
• 某轮之后 last-run.json 没更新：该轮可能不是 completed（被中断、报错、超长），或来自子代理会话，这两类都会跳过。
• 提炼失败：看 last-run.json 的 error 或 status.json 的 lastError；常见原因是当前路由不可用。
• 记忆太多/太杂：用 memory_forget 删掉；或直接编辑 memories.json 后重启。自动提炼的提示词已经要求「不记临时进度和可从代码直接读到的内容」。

开发

node --check index.js
node tests/memory.test.mjs

目录：

index.js            # 插件本体：事件、提炼、注入、三个工具
lib/memory.js       # 纯函数：归一化、相似度、解析、合并、裁剪、摘要
locale/{en,zh}.json # 插件管理里显示的名称与说明
icon.svg            # 插件图标
cordis.patch.yml    # bundle 的 Loader 插入行
tests/              # 无框架的 node 断言测试

无构建步骤、无第三方依赖；改完直接重新安装目录即可。

许可

MIT

---

English

Automatic long-term memory for DeepSeek Harness: after each finished turn it distills the few facts worth keeping across sessions, and feeds them back into later sessions.

Features

• Automatic distillation — when a turn completes, the user message and the assistant's final reply go to the same model route the session already uses, asking for at most three durable facts. No need to say "remember this".
• Global / workspace scopes — global holds who the user is and how they want answers (identity, preferences, taboos, language, style); workspace holds this directory's stack, commands, conventions, key decisions and pitfalls. They never mix.
• Deduplication — a new fact that closely restates an existing row is merged into it (timestamp refreshed, fuller wording kept) instead of piling up. At most 500 rows per store; the least recently updated are dropped first.
• Automatic recall — the first time a new session assembles its system prompt, a digest (≤ 1400 chars) of global plus current-workspace memories is injected. The digest is frozen for that session, so the system prompt never changes and prompt caching keeps working.
• Three tools — memory_search, memory_save, memory_forget for the cases automatic distillation cannot cover.
• Local and dependency-free — the store is one plain JSON file; no network calls of its own, no third-party npm packages.

Requirements

• DeepSeek Harness 0.2.x (verified on the desktop profile with 0.2.1-rc.3).
• Host-side only: no client code, no build step.

Install

1. Download this repository to a local directory (git clone, or download the ZIP and extract).
2. In DSH open Settings → Plugins, choose to install from a local directory, and point it at that folder. Scripted equivalent: call plugin_manager with install_bundle and pass the folder's absolute path as target.
3. Once installation reports application: applied, the plugin list shows Long-term Memory / 长期记忆.

There are no settings to configure. No restart is required; diagnostic files such as status.json exist from 0.1.1 on.

Usage

It is automatic. Just talk normally; the next turn or the next session knows what to remember and what to bring back.

Call the tools only when you want to be explicit:

Tool	What it does	Parameters
memory_search	Keyword search over memories; returns row ids	query (required), scope = auto/global/workspace, limit
memory_save	Save one durable fact; near-duplicates merge	text (required), scope = global/workspace, tags
memory_forget	Delete one row by id	id (required)

Detail panel — click the memory icon in the left sidebar to open a full page (the same page also sits in Settings). It reads the live store and analyses it: totals (all / global / workspace / new in 7 days), a health score (duplicate rows, rows untouched for 30 days and a failed last distillation each cost points and are listed as reasons), top tags, the last distillation's outcome and duration, and the full list with search, scope filtering and per-row delete.

Every row expands into "view conversation": the full user message and assistant reply of the turn it came from (with session id, turn number and time), plus the row's created/updated times, hit count, source and full workspace path. The source conversation is captured at distillation time; each row keeps its five most recent sources.

Daily diary — the panel has two tabs, Memory and Diary. The diary writes one entry per day from that day's memories: a short first-person note, a comment from the desk pet 「大肥鱼」, and the habits observed that day (habits are stored as global memories tagged habit, so they keep feeding later distillation and search). Entries live in diary.json; "rewrite today" regenerates the current one.

Choosing a scope:

• User-wide habits, name, language, taboos, taste → global.
• Things that only hold in this directory: language version, package manager, build commands, directory conventions, architecture decisions → workspace.

Where the data lives:

%DSH_HOME%\plugin-data\dsh-long-memory\      # Windows, default C:\Users\<you>\.dsh\...
~/.dsh/plugin-data/dsh-long-memory/          # macOS / Linux
├── memories.json    # the store (plain JSON, safe to inspect and edit)
├── last-run.json    # the latest distillation: added/merged counts, duration, error
└── status.json      # counters since boot: turns seen, successes, failures, last error

The DSH_HOME environment variable overrides the root. Deleting memories.json clears all memories.

Cost — one small extra request per completed turn (about 2k input tokens, up to 700 output), plus a ≤1400-character digest in the system prompt that is computed once per session, which keeps it cache-friendly.

Privacy — memories are distilled from the conversation and the request goes to the model route the session already uses; the store stays on your machine and the plugin uploads nothing.

How it works

1. It watches session/event: turn/start opens a buffer, which collects the text of user/message and assistant/message events.
2. On turn/end with reason = completed, it reads the current route from session.requestHeader() and asks the model for JSON in the form [{"scope":"global|workspace","text":"…","tags":["…"]}]. Subagent sessions (header.origin = subagent) and turns that did not complete are skipped. If the plugin was installed or reloaded mid-turn (missing the turn/start), it replays that turn from the session log instead.
3. Candidates are parsed, cleaned (≤ 300 chars each, at most 3) and compared with existing rows by containment similarity: ≥ 0.82 counts as the same fact and merges (keeping the longer wording); otherwise a new row is added.
4. When a new session first assembles its system prompt, global plus current-workspace rows are sorted by recency and injected as one digest. It does not change again during that session.
5. Similarity is computed over a token set of latin words and CJK character bigrams, taking the larger of Jaccard and containment — so "用户喜欢深色模式" and "用户偏好深色模式" are recognised as the same row.

Uninstall

Disable or remove dsh-long-memory in the plugin manager; the memory files stay and are reused if you re-enable it. To wipe everything, delete %DSH_HOME%\plugin-data\dsh-long-memory\.

Troubleshooting

• No digest in a new session — the digest is frozen per session; first check that memories.json actually has rows. An empty store injects nothing.
• last-run.json did not update after a turn — that turn probably was not completed (interrupted, errored, max-tokens) or came from a subagent session; both are skipped by design.
• Distillation failed — read error in last-run.json or lastError in status.json; the usual cause is an unavailable model route.
• Too many or noisy memories — delete them with memory_forget, or edit memories.json and restart. The extraction prompt already asks the model to skip transient progress and anything obvious from the code.

Development

node --check index.js
node tests/memory.test.mjs

Layout:

index.js            # the plugin: events, distillation, prompt digest, three tools
lib/memory.js       # pure functions: normalisation, similarity, parsing, merge, prune, brief
locale/{en,zh}.json # display name and description in the plugin manager
icon.svg            # plugin icon
cordis.patch.yml    # the bundle's Loader insert row
tests/              # dependency-free node assertion tests

No build step, no third-party dependencies; reinstall the directory after editing.

License

MIT