---

这是什么

Daemonkey 不是又一个网页聊天框。它是装在你电脑上的后台程序：第一次对话之后，你就有了属于自己的专属 AI。换模型、换电脑，它怎么叫、怎么叫你、聊过什么还在。

别人记住的是关于你的几条资料。
这里留下的是你们一起做过的事，加上它自己也在长。

换模型没关系。属于你的那一份还在。

_{桌宠：贴在屏幕边上，待机眨眼、干活跟着忙、活干完邀功、累了趴下 —— 状态跟着你们正在干的事变。}

_{用量面板：每次开口用了多少 token、走的哪个模型、缓存命中多少，全记在本地。}

---

记忆不是人设，是两套治理

别人的陪伴，是一句「你要温柔一点」的提示词。Daemonkey 的性格，是你们一起经历的事长出来的 —— 而且这套记忆，正是它能把活干好的同一个底子。

两套分开存、分开管，谁也不会挤掉谁：

两套都写在你自己的硬盘上 · 记事本能打开、能改、能带走。

_{写入 → 两套治理 → 取出 → 治理：一条记忆从进来到被用上的全程。}

_{记忆星图：手册和记忆聚成星系。灵魂来自记忆，不是写在提示词里的人设。}

---

打开之后能干什么

---

为什么越用越便宜、越记得住

每次开口，模型都要先读一段固定说明书（提示词前缀），再读你刚说的话。说明书越短、越稳定，后面几句就越不用把前面那一大段再付一遍钱。

出厂实测（2026-09，纯净版、空画像）：

怎么做到的：每轮会变的内容（时间、进度）放到最后，前面保持不动，磁盘缓存就能对上。常用的工具全文写进说明书，其余只留名字，用到了再展开。

说明书不是死的。 装配台把 41 个条目、141 件工具摊开：写代码就装写代码那几段，写作就装写作那几段 —— 左边挑前缀条目、中间挑工具、右边实时看拼出来多少 token。装得薄，本机跑小模型时上下文也更宽裕。

画像用久了会分层：改说话的短条每轮都在，流水不必每次灌完 —— 实测每轮常驻说明大约少 68%。

_{装配台：左边挑前缀条目 · 中间挑工具 · 右边实时看拼出来多少 token。}

记得多不是本事，记得干净才是。先挡住不该记的；找旧事时先全文搜，再用模型挑真相关的。你拒过的下次会排到前面，做过的下次更知道怎么做。

---

ATM-Bench：同一个模型，分差在架构

ATM-Bench（arXiv 2603.01990）测的是长期记忆问答：四年跨度的相册、视频、邮件，31 道难题。我们把 Daemonkey 自己的记忆引擎接进去跑，答题和打分都用 DeepSeek 官方 deepseek-v4-flash，不拿别人的宣传图。

---

它怎么跑起来

没有云数据库。后台程序在你这台机器上，旁边接 OpenAI 兼容 API，底下是文件。

你
 ├── 启动器（Windows 双击 exe / Mac 一条命令）
 └── 本机 Daemonkey
       ├── 大脑：任何 OpenAI 兼容接口（官方或中转，模型你自己选）
       ├── 工作台 · 房间 · 微信 / 飞书
       └── 只存在你硬盘上的东西
             它是谁、你是谁、对话、稿、记忆索引

工作台、房间、微信是外壳。换外壳，里面还是同一个它。

---

和普通聊天框差在哪

---

三步上手

Windows

1. 双击 Daemonkey.exe → 环境 → 开始安装。第一次大约一分钟。
2. 回到启动页，点启动。浏览器会自己打开。
3. 填一个大模型 API Key，给它起名字，告诉它怎么叫你。这些写进画像。

macOS / Linux

chmod +x start.sh && ./start.sh

Mac 也可以把 Daemonkey.app 放进这份代码的根目录（和 tools/ 同级）再双击。不要只把应用单独丢进「应用程序」。详见 MAC-GUIDE.md。

你需要：Python 3.10+（安装时勾选 Add to PATH），以及一个 OpenAI 兼容 API 的 Key（官方或中转都行）。

崩了也不慌：

ZIP 包用户：启动器第一次会配好官方升级源（Gitee 主、GitHub 备份）。之后在启动器里「检查更新」，或对话里说一声即可。

---

下一步

完整路线图见 ROADMAP.md。一句话：近处把工作台和房间打磨扎实；远处是多设备还是同一个它，以及真正的桌面机器人。不做云端 SaaS，不绑死一家模型。

---

许可

Copyright © 2026 vaan21th · AGPL-3.0。

可以自用、修改、分发。改过的版本——哪怕只是架成网上服务给别人用——也要按同一协议公开源码。

永久免费。 有人跟你收费，去找卖家退款。官方只在 B站 / 抖音发布。

欢迎提 Issue、提 PR。新本事请合回内核，这样每个人的搭档都能用上。

---

What it is

Daemonkey is a daemon on your own machine, not another browser chat box. After the first conversation you have your own dedicated AI. Swap the model or the PC — the profile and diary stay.

Others keep a few facts about you.
Here what stays is the work you did together, plus its own growth.

Swap the model. The dedicated one stays.

_{The desktop pet: sits by the screen edge — blinks when idle, busy when you work, celebrates when done, lies down when tired. Its state follows what you're doing.}

_{Usage panel: tokens per turn, which model ran, cache hit rate — all recorded locally.}

---

Memory is not a persona — it's two governed tracks

Other companions are a prompt that says “be gentler”. Daemonkey's character grew out of things you went through together — and that same memory is the very foundation that makes it good at the work.

Two tracks, stored and governed separately, so neither crowds out the other:

Both tracks live on your own disk · readable, editable, portable.

_{Write → two tracks → read → govern: the whole trip of one memory, from arrival to use.}

_{Memory star map: playbooks cluster into galaxies. The companion is memory, not a persona prompt.}

---

What you can do

---

Cheaper over a long chat, and it actually remembers

Every turn the model reads a fixed prefix (the standing instructions), then your new line. Shorter and more stable prefix = later turns do not pay for the whole booklet again.

Factory measurement (2026-09, clean build, empty profile):

How: volatile bits (time, progress) go at the end; the front stays still, so disk cache matches. Everyday tools stay in the manual; the rest keep only their names, expanded when used.

The manual isn't fixed. The Rig lays out 41 entries and 141 tools: writing code loads the coding parts, writing prose loads the prose parts — pick entries on the left, tools in the middle, live token count on the right. Assemble it thin and a small local model gets more room to breathe.

A long-used profile is layered: the short card every turn, the diary on demand — about 68% less standing text per turn in our measurement.

_{The Rig: prefix entries on the left · tools in the middle · live token count on the right.}

---

ATM-Bench: same model, the gap is architecture

ATM-Bench (arXiv 2603.01990) is long-horizon memory QA: four years of photos, video, mail; 31 hard items. We plug in Daemonkey’s own memory engine. Answers and the judge use official DeepSeek deepseek-v4-flash.

---

How it runs

No cloud database. Daemonkey runs on your machine, talks to an OpenAI-compatible API, and keeps files on disk.

you
 ├── launcher (Windows exe / macOS one command)
 └── local Daemonkey
       ├── brain: any OpenAI-compatible endpoint (official or proxy)
       ├── workbench · room · WeChat / Feishu
       └── only on your disk
             who it is, who you are, chats, files, memory index

Workbench, room, and chat apps are shells. New shell, same companion.

---

Versus a chat box

---

Quick start

Windows: double-click Daemonkey.exe → Environment → Install (~1 min) → Start → paste an LLM API key, name it, say how to address you.

macOS / Linux:

chmod +x start.sh && ./start.sh

On a Mac you can also put Daemonkey.app in the project root (next to tools/) and double-click. Do not drop the app into Applications by itself. See MAC-GUIDE.md.

You need Python 3.10+ (tick Add to PATH) and an OpenAI-compatible API key.

repair.bat · ROLLBACK.bat · verify.bat if something breaks. ZIP installs get Gitee + GitHub update sources on first launch.

---

License

Copyright © 2026 vaan21th. AGPL-3.0. Free forever. If someone charged you, ask them for a refund. Official posts: Bilibili / Douyin.

Issues and PRs welcome. New abilities should land in the kernel so every companion gets them.