dsh-lite-memory

dsh-lite-memory

没有常驻的记忆正文：不写系统提示，不逐轮注入，命中才读。 No memory in the prompt: nothing injected per turn, read only on a hit.

---

它是什么 / What it is

给 DeepSeek Harness 的轻量记忆：不写常驻系统提示、不逐轮注入、不扫描文件。记忆只在两种时刻出现——你说「记一下」，或者这句话跟已有条目真有关（最多三条标题，正文仍按需读）。

Lightweight memory for the DeepSeek Harness. No standing system prompt, nothing injected per turn, no file scanning. Memory appears at exactly two moments: you ask it to remember, or the message genuinely overlaps an existing entry (at most three titles; bodies are read on demand).

触发方式 / Triggers

场景	行为
L0 你明确要求「记住这个」	立即写盘，一条确认行
L1 它发现有跨会话价值的经验	一行灰字 memory · suggested "X" · sug_x + 一个 keep 按钮；标题加一句话，看得懂才点，不点就永远不写，48 小时内不重复
L2 常识、一次性参数、凭据的值	永不写入

这些行只属于产生它的那一轮：每轮按那轮自己的工具结果重建，重进会话还能看见，别的轮不会混进来。再记一次时插件自己判是更新还是新增：同层、同前置条件下，标题（或这条用过的旧标题）完全一致就更新原条目，换过说法而标题变了的旧标题记成别名以后也认得。像但不同名的不写：先在结果里回「相近：X」和它的 id，确认要更新就带那个 id 再存，确认是另一条才加 resolution="distinct"。更新时没给的字段保持原值，正文整段替换——旧正文只作内部撤回依据，不展示、不参与检索；前置条件不同的同一主题是两条。

写入时分两层：默认只在这个项目生效；用户说「以后所有项目都这样」才进全局层（200 条起提示，硬上限 1000 条），检索时一起查。同一主题在不同前提下用 when 区分，各存一条。

写入前先蒸馏：标题是结论，正文只留原则、判断规则和触发条件，不把具体尺寸、字号、层级、颜色或一次性数字记进来；需要这些值时按记忆里的指针查当前代码或文档。单条正文超过 4096 字节会被拒绝，先删细节再存；摘要超长自动截到 60 字。

触发词：记住、记一下、上次、规范、约定、路径、凭据怎么取。

真命中看这句话和记忆的标题（含旧标题）、标签、触发词、摘要有没有对得上的词：有就多一行「已有记忆可能相关」，最多三条标题和 id，每条前面标归属（[global] 或项目目录名如 [DSH]）；没对上就不出现，不改命中计数，正文要 memory_read 才读。一条记忆写了触发词，用户说到那个词也算命中。

Case	Behaviour
L0 you ask it to remember	written immediately, one confirmation line
L1 it spots something worth keeping	one muted line memory · suggested "X" · sug_x with a single keep button — readable before you click, nothing written unless you click, at most once per two days
L2 common sense, one-off parameters, credential values	never written

Those lines belong to the turn that made them: each turn is rebuilt from its own tool results, so reopening the session shows them again and no other turn leaks in. Saving again updates an entry only when the title (or one of its earlier titles) matches exactly under the same precondition. A merely similar title writes nothing first: the result reports the closest entry and its id, so the caller either updates that id or re-saves with resolution="distinct". An update keeps every field it was not given and replaces the body wholesale; the previous body is kept only for internal undo and never shown or searched. The same subject under different preconditions stays two entries.

Writes land in one of two layers: the current project by default, or the global layer (warns past 200, hard ceiling 1000) when the user says the rule holds for every project. Both layers are searched together, and a when field separates the same subject under different preconditions.

Before writing, distill: the title is the conclusion and the body keeps only the principle, decision rule and trigger condition. Concrete dimensions, font sizes, z-index, colors and one-off numbers stay out; the memory points at where to read the current value instead. A body over 4096 bytes is refused, so drop detail before saving; an over-long summary is truncated to 60 characters.

点一下就存 / One click, no turn

keep 是同源直写：不产生消息、不动输入框、不打断当前回复。按钮变 kept 就结束了。真写过的行尾才有 ✕，按 id 撤销：改过的条目还原成上一版（显示 restored），刚建的条目移入归档（显示 archived）；条目后来又改过就停下并说明已经变过，不覆盖后来的修改。没变、需要确认、保存失败的行没有撤销。

keep writes straight to disk from the page: no message, no extra model turn, the input box untouched. The button turns into kept and that is the whole interaction. Only a row that really changed gets the ✕, and it undoes by id: an edited entry goes back to its previous body (restored), a freshly created one moves to the archive (archived); if the entry changed again it stops and says so instead of overwriting. Noop, needs-confirmation and failed saves get no undo.

每轮引用条 / Per-turn references

每次回复末尾、在「已编辑 N 个文件」下面，按这一轮实际发生的工具结果留几行：本轮参考 N 条 和 本轮更新 N 条。点开是记忆名，再点一条就地展开当前正文；名字和正文都取当前值——改过名显示新名字，归档的标「已归档」，删掉的显示「已删除」。同一轮同一条只出现一次；没读过也没写过就整段不出现。切换会话不丢，每轮只按自己的工具结果重建。

Each completed turn keeps a few lines under 「已编辑 N 个文件」, built from that turn's own tool results: 本轮参考 N 条 and 本轮更新 N 条. A line expands to memory names, a name expands to the body as it is now — current title, current body, archived marked, deleted said outright. One row per entry per turn; nothing read or written, nothing shown. Switching conversations keeps them, because each turn is rebuilt from its own tool results only.

记忆面板 / Memory panel

会话头的大脑图标打开居中的记忆面板。顶部是「启用记忆」「自进化」两个开关，右上角「显示设置」里另有「启用图谱」「启用记忆宫殿」两个只影响展示的开关，默认关闭，按项目记住；内容分「记忆」「自进化」两个页签。默认看当前项目，也可以切到其他项目或全局层；切过去只读，不回写、不迁移旧格式。

「记忆」页签里，搜索框右侧可以在「列表」「图谱」「宫殿」之间切换；列表、图谱或宫殿占满内容宽度，选中记忆后右侧才出现正文与归属。图谱按真实数据画，点记忆节点读正文，点标签或主题节点转成搜索，右上角可缩放。宫殿按类型分四间：习惯与规范、方法与经验、资料与入口、其他记录；一次只展开一间，每间每页 24 条，点展品只读正文。宫殿只决定怎么陈列，不改记忆、学习状态和额度。「自进化」页签只作用于当前项目，按「全部 / 候选 / 试用 / 生效 / 暂停」筛选，也能切列表与图谱；详情给出当前经验、适用范围、当前状态、学习依据和变化记录。候选可采纳、修改、放弃，试用可采纳、修改、暂停，生效可修改、暂停，暂停可恢复或查看记录。

The brain icon in the session header opens a centered memory panel. Memory and Self-evolution sit at the top, and Display settings at the top right add two presentation-only switches, Graph and Memory palace, both off by default and remembered per project. Two tabs, Memories and Self-evolution, hold the content. It starts on the current project and can switch to another project or the global layer; switched scopes are read-only and never migrate or create state.

On the Memories tab, the search row switches between list, graph and palace; either view fills the content width, and the body pane only appears after a memory is selected. The graph is drawn from real data: a memory node opens its body, a tag or topic hub becomes a search, and the top-right controls zoom it. The palace sorts memories into four rooms by kind: Habits and conventions, Methods and experience, References and entries, and Other notes; one room opens at a time, each page holds 24 items, and an exhibit only opens its body. The palace changes presentation only, never memory, learning state or quota. The Self-evolution tab acts on the current project only, filters by All / Candidate / Trial / Active / Paused, and also offers list and graph views. It shows the current rule, scope, state, evidence and change history. Candidates can be adopted, edited or discarded; trials can be adopted, edited or paused; active rules can be edited or paused; paused rules can be resumed or inspected.

• 关闭项目记忆后，这个项目不再使用任何记忆（包括全局层）：自动注入、检索、读取、引用、保存全部停，服务端也拒绝记忆工具和面板修改动作。自进化同时关闭并禁用；重新打开记忆后自进化仍为关闭，要单独开启。

• 关闭只改权限，不删数据：候选、试用、生效、证据、整理额度和默认绑定都保留，计数照常显示。

• 记忆工具与面板管理动作都在服务端拒绝，不能靠改前端绕过；其他项目不受影响。

• 范围只来自服务端枚举的层；不接受客户端拼路径，也不能写别的项目。

• 图谱只按数据画：每条有效记忆一个节点，共享标签和主题做枢纽，显式关联画实线；自进化图谱只连共用条件。位置由 id 决定，同一批数据每次一样。

• Turning project memory off stops all memory use in that project, including the global layer: no injection, search, read, citation or save, and the server rejects both memory tools and panel mutations. Self-evolution turns off with it; after memory is reopened, self-evolution stays off until switched on.

• Closing is a permission change, not deletion: candidates, trials, active rules, evidence, organize quota and default bindings are kept, and counts stay visible.

• Memory tools and panel management actions are rejected server-side, so a modified client cannot bypass the switch; other projects are unaffected.

• Scopes come only from the server-side list; the page cannot assemble a path or write another project's notes.

• The graph is data only: one node per active memory, shared tags and topics as hubs, explicit links as solid edges; positions derive from ids, so the same data looks the same every time.

为什么轻 / Why it stays light

• 0 依赖，存储层无定时器、无监听、无网络，只有同步单文件写；写入用一把有超时的锁，别的进程正在写就明确报「忙」，不抢别人的锁。

• 搜索最多 4 条、400 字；正文只在 memory_read 里出现，全文永不进上下文。

• 普通检索只列当前有效的记录；已经归档或被更新替换的旧说法不混进来，要查历史得在 memory_search 里显式要 history。

• 上限：单条正文 4096 字节、整库正文 1 MB、每层 1000 条、标题与摘要各 60 字；每层 200 条只当提示线。90 天未用只留 id + 标题。到了硬上限不拒绝新记忆：把同层最久没被读的一条移进归档，用户固定过的规则不动，保存结果里写明是哪条（正文与完整记录都留着，可人工取回），单条本身超预算才报错；用户明确说「别记」也进归档。文件从不删除。

• 一个项目一个目录，项目之间互相看不见。

• Zero dependencies. The store has no timers, no listeners, no network — just synchronous single-file writes, guarded by a lock that gives up after 500 ms instead of stealing it.

• Search returns at most 4 hits and 400 characters. Full text only appears in memory_read.

• Normal search lists current entries only; archived notes and superseded wording stay out unless memory_search asks for history.

• Limits: 4096 bytes per body, 1 MB of bodies per store, 1000 entries per layer, 60 characters per title or summary; 200 entries per layer is just a warning line. At the hard ceiling a save is never refused: the least recently read entry in the same layer moves to archive (pinned rules are never evicted) and the result names it (body and full record kept, restorable by hand); only an entry that alone exceeds the budget fails. Saying 「别记」 also archives. Files are never deleted.

• One directory per project, so projects never see each other's notes.

项目自进化〔实验〕 / Project self-evolution (experimental)

默认关闭，只对当前项目生效。开启后，它从你的纠正和认可里整理出少量项目规范：纠正更新做法，认可作品只给已有规范留一条带维度的样例；执行漏掉只补检查点，不改写成禁令。下次匹配任务按需读取，不逐轮注入。会增加一次有上限的后台整理调用，以及命中的记忆读取消耗。

Off by default, per project only. When enabled it distills a few project rules from your corrections and approvals: corrections revise the approach, approvals attach one bounded sample to an existing rule, and an execution miss adds a checkpoint instead of a ban. Matching tasks read them on demand, never per turn. It adds one bounded background organize call and the memory reads that actually hit.

• 一次整理最多 2 处改动、768 输出 token、15 秒；输入不超过 8 KiB，同一作品最多 2 次，同一项目 24 小时内最多 6 次。

• 成果停在 pending / trial / active / paused；明确说「以后都这样」或两个独立作品支持才 active，试用与暂停 30 天后清理。试用会参与条件匹配任务的读取，不是预览稿；到期自动暂停。

• 规则 32 条、待定 12 条、每规则样例 4 个/检查点 3 个/来源 3 条、旧版本 2 个，都有硬上限；到顶先暂停新学习，不静默丢已生效规则。

• 面板显示整理次数和已知用量；usage 没返回就写未知，不补零。

• 关闭后停止学习、读取和晚到写入；成果保留可查看。记忆总开关关闭时自进化一定也是关闭，不能单独打开。

• 同一项目换栏目、新一期、临时例外不会互相继承；仓库子目录、符号链接、worktree 视为同一项目，不同克隆分开。

• At most 2 changes, 768 output tokens and 15 s per organize call; input ≤ 8 KiB, ≤ 2 calls per episode and ≤ 6 per project in 24 h.

• Results stay pending / trial / active / paused; they become active only on an explicit long-term ask or two independent episodes; trial and paused entries expire after 30 days. A trial is read by matching tasks — it is not a preview — and pauses when it expires.

• Hard caps: 32 rules, 12 pending, 4 samples / 3 checkpoints / 3 sources per rule, 2 old versions. At the ceiling new learning pauses; active explicit rules are never silently dropped.

• The panel reports organize calls and known usage; missing usage is shown as unknown, never zero.

• Turning it off stops learning, reads and late writes; results stay viewable. While project memory is off, self-evolution is necessarily off and cannot be switched on by itself.

• Switching columns, a new episode and one-off exceptions do not leak into each other; a repository subdirectory, symlink or worktree is the same project, separate clones are not.

安装 / Install

Plugin Manager 里搜 dsh-lite-memory 直接装；或 npm pack 后用 Plugin Manager 安装生成的 .tgz。

Install from the DSH Plugin Manager (search dsh-lite-memory), or npm pack and install the resulting .tgz through it.

工具 / Tools

memory_save · memory_suggest · memory_search · memory_read · memory_forget

完整触发规则见 TRIGGERS.md；性能与 UI 门禁的实测数字见 docs/BENCH.md。 / Full trigger rules: TRIGGERS.md; measured performance and UI-gate evidence: docs/BENCH.md.

License

MIT