dsh-memory-system-prompt

dsh-memory-system-prompt

中文 | English

全局记忆插件：AI 可以主动把事实写进记忆文件，并在后续对话里按需读取。记忆索引每轮自动注入系统提示，不占用对话上下文。

同时包含用户级系统提示词功能（原先独立的 dsh-system-prompt 已并入本包），两部分共用同一个设置页。

它解决什么

模型本身没有跨会话记忆。这个插件给出一份可持久、可检索的记忆区：

• AI 主动写：对话中发现的稳定事实（偏好、约定、环境细节）由模型调用工具写入；
• 索引常驻：每轮把「标题 + 一句话摘要」注入系统提示，模型知道有什么；
• 正文按需：需要细节时再调工具读全文，不把整个记忆库塞进上下文；
• 可按对话关：某个对话不想带记忆时，单独关掉。

工具

工具	作用
memory_save	保存或更新一条记忆（同名覆盖）
memory_read	按标识读取一条记忆的正文
memory_list	列出全部记忆（默认只返回索引）
memory_delete	删除一条记忆

用户级系统提示词

写一次，对所有对话生效；也可以按对话单独禁用。

• 自定义正文：任何想长期强加给模型的要求（语气、格式、禁止事项）；
• 九维风格档位：不想写长文时，用档位拼出一段行为指令（基础语气、亲切程度、热情程度、表情符号、可视化回答、标题和列表、快速回答、提示词建议、回答长度）；
• 两种注入位置：作为系统提示的一部分，或作为上下文片段。

档位默认都是「默认」，全部保持默认时不产生任何额外内容。

界面

• 输入框左侧开关：三态（关 / 部分 / 开），按对话生效；
• 设置 → 记忆与提示词：管理全局开关、单条记忆、系统提示词与风格档位、首次使用引导。

存储

路径	内容
~/.dsh/memory/	每条记忆一个文件
~/.dsh/memory-settings.json	记忆的全局开关与偏好
~/.dsh/system-prompt/settings.json	系统提示词与风格档位

安装

推荐：内置插件安装程序。 打开「设置 → 内置插件」，粘贴包名，点安装，再重启应用：

dsh-memory-system-prompt

安装程序会自动写进 dsh.profile.bundles，不用手改任何文件。

命令行方式（先完全退出应用）：

cd ~/.dsh/profiles/desktop
pnpm add dsh-memory-system-prompt

然后把包名加进 dsh.profile.bundles，重启应用。

注意

• 记忆索引通过 systemPrompt.context 注入，不是 agent.inject——后者会排队成一条用户消息，导致模型在回答完再自问自答一轮。
• 索引刷新挂在 agent/created 和 turn/start 上，因为在 assemble 之后才拿得到最新内容。
• 系统提示词部分不把 systemPrompt 写进 inject，而是运行时用 agent.ctx.get('systemPrompt') 取。硬注入会让 fiber 在服务未就绪时被挂起，表现是插件静默不加载、设置里一片空白。

---

English | 中文

Global memory for DeepSeek Harness: the model can write facts into a persistent store and read them back later. The index is injected into the system prompt every turn, without consuming conversation context.

The package also carries the user-level system prompt feature (the former standalone dsh-system-prompt is merged in). Both halves share one settings page.

What it solves

Models have no cross-session memory. This plugin provides a persistent, searchable memory area:

• The model writes: stable facts discovered in conversation (preferences, conventions, environment details) are saved through tools;
• The index is always present: each turn injects "title + one-line summary" into the system prompt, so the model knows what exists;
• Bodies are on demand: the full text is read only when needed, so the whole store never floods the context;
• Per-conversation opt-out: a conversation that should not carry memory can turn it off.

Tools

Tool	Purpose
memory_save	Save or update one memory (same slug overwrites)
memory_read	Read one memory body by slug
memory_list	List memories (index only by default)
memory_delete	Delete one memory

User-level system prompt

Written once, applied to every conversation; can also be disabled per conversation.

• Custom body: anything you want to impose on the model long-term (tone, format, prohibitions);
• Nine style dials: build a behavior instruction from dials instead of prose — base tone, warmth, enthusiasm, emoji, visual answers, headings and lists, quick answers, suggestions, verbosity;
• Two injection positions: part of the system prompt, or a context fragment.

Every dial defaults to "default"; keeping all of them default produces no extra content.

UI

• Composer toggle (left side): three states (off / partial / on), per conversation;
• Settings → Memory & Prompt: global switch, individual memories, system prompt and style dials, first-run guide.

Storage

Path	Contents
~/.dsh/memory/	One file per memory
~/.dsh/memory-settings.json	Global switch and preferences for memory
~/.dsh/system-prompt/settings.json	System prompt and style dials

Install

The built-in plugin installer is the recommended way. Open Settings → Built-in Plugins, paste the package name, press install, then restart the app:

dsh-memory-system-prompt

The installer writes the entry into dsh.profile.bundles for you; no manual file edits.

From the command line (quit the app completely first):

cd ~/.dsh/profiles/desktop
pnpm add dsh-memory-system-prompt

Add the package name to dsh.profile.bundles and restart the app.

Notes

• The memory index is injected through systemPrompt.context, not agent.inject. The latter queues a user message, which makes the model answer its own question for another turn.
• The index refreshes on agent/created and turn/start, because the content is only current after assembly.
• The system-prompt half deliberately does not list systemPrompt in inject; it resolves the service at runtime via agent.ctx.get('systemPrompt'). A hard injection parks the fiber while the service is not ready, which shows up as a silently unloaded plugin and an empty settings page.