By day it records what happens here. At night it reads the record back: which of it is habit, which is coincidence. Only what holds up changes its mind. You write no rules for any of this. 白天它记下这里发生的事,夜里从头过一遍:哪些是习惯、哪些是巧合,分清楚了才更新。规则一条都不用你写。
Automation platforms hand you an editor and wish you luck. Here the division of labor is reversed: going about your day is your job, noticing the routine is its job. 自动化平台把编辑器递给你,剩下靠你自己。这里分工反过来:该干嘛干嘛是你的事,看出规律是它的事。
Every condition is yours to think of, and yours to fix when life changes. The rule does not notice that it went stale. You do, usually late. 每个条件靠你想到,生活一变还得靠你来修。规则不会发现自己过时了。发现的是你,而且通常晚了。
You write zero rules. It brings the observed routine to you as a question, and nothing becomes standing behavior without your yes. When life changes, it learns the routine over again the same way. 你一条规则都不用写。它把看出来的规律变成一个问题拿给你,你点头之前,什么都不会变成惯例。生活变了,规律照同样的路子重新学。
Record, learn, propose, feedback, and the feedback flows back into what it believes. Your verdicts are training data: every yes and no you give makes the next proposal more accurate. 记录、学习、建议、反馈,反馈再回流进它的认识。你的每一次「好」和「不用」都是训练材料,下一次提议因此更准。
Behind this cycle run four agents with separate jobs: one chats, one tidies the space file, one wakes on schedule, one learns at night. Up front you only ever face one butler. 这一圈背后是四个各管一摊的 Agent:一个管对话、一个管整理空间档案、一个到点醒来跑任务、一个夜里学习。台前,你永远只面对一个管家。
Learning that proposes actions in your space has to be strict about evidence. These are the checks every candidate routine passes before it ever reaches you. 敢在你的空间里提议动作,证据就得经得起挑。每条规律送到你面前之前,先过这几道。
Accepting a card is one tap, and the action it triggers stays undoable inside a short window. Declining teaches it too: say no and that routine's standing drops at once. 同意就一下,触发的动作在反悔窗内都能撤回。拒绝同样是教它:你说一次「不用」,那条规律的分量立刻降下去。