Muse？Claws！Muse?Claws!

每换一个平台，就得重新自我介绍一次Every new platform, another self-introduction

你是不是也这样：根据需求换着用不同的 AI，每家都存了一份关于你的记忆，互相却不通。

Sound familiar? You switch between AIs for different jobs, and each one keeps its own memory of you, but they never talk to each other.

早上跟 ChatGPT 说早餐换成燕麦了，晚上问 Claude 明天吃什么，它一无所知。

In the morning you tell ChatGPT you've switched to oats for breakfast. In the evening you ask Claude what to eat tomorrow, and it has no idea.

Muse 把这件事做到了头：模型是它家的，记忆从零开始，数据放在它的云里，别的 AI 无法实时读取。拆开看，它就是一只 claw 套着 app 的壳，接了一堆 MCP，读着你早就存好的数据。问题在于太中心化。

Muse takes this to the extreme: the model is Meta's, the memory starts from zero, and your data sits in their cloud, where no other AI can read it live. Take it apart and it's a claw in an app shell, plugged into a pile of MCPs and reading data you'd already handed over. The problem is that it's centralized.

手机为什么人人都在用？因为它不绑死你Why does everyone use a phone? It doesn't lock you in

买安卓还是苹果都行，没有哪家逼你只用自家的软件；无数创作者做的 app，拼成了你每天的生活。下一代设备起码也得这样：

Android or iPhone, neither forces you to use only its own apps, and apps built by countless creators make up your day. The next device should do at least as well:

模型自己挑，记忆带着走，功能按需求自己装、自己做。

Pick your own model, take your memory with you, and install or build the features you need.

模型可以换，功能可以换，但记忆只有一棵树Models can change, features can change, but there's only one memory tree

所以第一步先解决记忆。世界树是跑在你自己服务器上的一个 MCP 服务：树干是从 ChatGPT、Claude、Gemini 的导出里提炼出来的你；Claude.ai、ChatGPT、Gemini、Notion 和你的 claw 都是树上的枝，开口前先读树，学到关于你的新事实就写回去，每条都带来源和日期。

So memory comes first. The memory tree is an MCP server on your own machine. Its trunk is you, distilled from your ChatGPT, Claude and Gemini exports. Claude.ai, ChatGPT, Gemini, Notion and your claw are its branches: they read the tree before answering and write back new facts about you, each stamped with its source and date.

它们会自己动起来They move on their own

有了记忆，agent 才能主动。下面是作者 9 月 24 日真实的一天，每一条都是 agent 自己推过来的，没人去问。

With memory in place, agents can speak first. Here is a real day from the author's life, 24 September: every message below was pushed by an agent, unprompted.

1. 08:00
   起床：看昨晚睡得怎样，出今天的训练建议Wake-up: checks last night's sleep, writes today's training plan今天练 Pull B，按计划来，重量先不加Pull B today, as planned. Keep the weights where they are.
2. 13:30
   练前 90 分钟：现在该吃什么90 minutes before training: what to eat now现在吃 1 根香蕉，练完 30 分钟内冲蛋白粉A banana now, and a protein shake within 30 minutes after.
3. 16:21
   练完：训练软件一存，自动复盘Workout saved in the training app: an automatic review17 组做了 14 组，今晚 23:30 前睡，明天休息14 of 17 sets done. In bed before 23:30 tonight, rest day tomorrow.
4. 16:22
   练后餐：按今天已经吃的热量来配Post-workout meal, based on what's been eaten today蛋白粉 20 g、香蕉 1 根、开心果 25 gProtein powder 20 g, one banana, pistachios 25 g.
5. 19:30
   晚饭：按剩下的额度算好份量Dinner, portioned to what's left of the day's budget意面 80 g（生重）配西冷 80 g，蔬菜随意Pasta 80 g (dry) with sirloin 80 g, vegetables as you like.

上面是 app 的截图，用的是演示数据。

The app screenshots above use demo data.

作者交给 agent 的五件事Five things the author handed to agents

作者用两天时间，让 Fable 帮忙做了出来。

The author built them in two days, with Fable's help.

• 健身Fitness训练软件有 MCP，agent 实时看到练了什么：前一晚睡得好就照常练，练完按当天吃过的热量配练后餐。The training app has an MCP server, so the agent sees each workout as it happens. After a good night's sleep, training goes ahead as planned; afterwards it suggests a post-workout meal from what's been eaten that day.
• 饮食Diet读记下的每一餐，摸清饮食习惯，帮忙备餐，也给其他 agent 提供数据。It reads every logged meal, learns the eating habits, helps plan meals and shares the data with the other agents.
• 健康HealthApple Watch 的数据经 Apple 健康给到 agent，分析睡眠，留意心率。Apple Watch data reaches the agent through Apple Health. It analyzes sleep and keeps an eye on heart rate.
• 申请硕士Grad school接上邮箱和 Google 表格，盯申请进度，提醒 ddl。Connected to email and Google Sheets, it tracks every application and flags deadlines.
• 求职Job hunting给它 CV 和 cover letter，之后每个岗位只要丢一份 JD，它来改一版。Given the CV and cover letter once, it tailors a version for each job description it's handed.

这五个 agent 正在整理成功能包，放进仓库。

These five agents are being turned into packs for the repo.

你的需求，也会有人需要What you need, someone else needs too

每个人的需求不一样，但你的需求，这世界上总有人也有。手机因为有了做 app 的人才变得好用，agent 也一样：缺的是创作者，不是新的账单。

Everyone's needs are different, but whatever you need, someone out there needs it too. Phones got good because people built apps for them, and agents are the same: what's missing is creators, not another bill.

Muse 和自己搭，差在哪Muse vs. building your own

除了 Meta 自带的用户基数，Muse 在技术上没有护城河：每家模型厂商都能很快做出同样的东西，就像当年手机厂商互相竞争。

Apart from Meta's user base, Muse has no technical moat: any model company could ship the same thing fast, just like the phone makers' race years ago.

* OpenAI、Anthropic、Kimi 的订阅现在都能给 claw 用，不走 API 计费。各家条款不同，请自行确认。

* OpenAI, Anthropic and Kimi subscriptions can currently run a claw instead of API billing. Terms differ by provider, so check yours.

它是怎么搭起来的How it fits together

Meta 把 agent 跑在自家服务器上，再套一个 app。OpenMousse 也套这个壳，但 agent 跑在你的机器上，用的是你自己的模型订阅。

Meta runs the agent on its own servers and wraps it in an app. OpenMousse wraps the same kind of shell, but the agent runs on your machine, on your own model plan.