我常用的几个 AI 都不了解我：不知道我昨晚睡得怎样、这周有哪些截止日期、我在意什么，每次都得从头解释。所以我把自己的生活记录和想法整理成一套系统，每天自动更新，AI 回答之前先查它。The AI assistants I use every day knew nothing about me: how I slept, what was due this week, what I care about. I had to explain myself every time. So I organised my daily records and my own thinking into one system that updates itself, and the assistants check it before they answer.
Apple 健康里 2017 年以来的睡眠、心率和步数，MacBook、iPhone 和 iPad 的屏幕时间，看过的视频，Canvas 上的课程截止日期。大多每 2 小时自动更新一次，存在我家里的 Mac mini 上。Sleep, heart rate and steps from Apple Health going back to 2017; screen time on my MacBook, iPhone and iPad; videos I have watched; course deadlines from Canvas. Most of it refreshes every two hours and is stored on a Mac mini at home.
昨晚睡得怎样，昨天的时间花在做事还是娱乐，AI 帮我跑了多长时间的任务（按 Claude Code 和 Codex 的本地记录算），今天有哪些课、截止日期和训练，再对照我定的目标，提醒今天先调整什么。哪一路数据断了，它会直说停在哪天。How I slept, whether yesterday went on work or entertainment, how long AI spent on tasks for me (from Claude Code and Codex logs), today's classes, deadlines and training, and what to adjust first against the goals I set. If a data source has stopped, it says which day the data ends.
Claude、Codex 和 ChatGPT，还有在 Telegram 上和我聊天的现成助手 Hermes，都先读同一份说明文件：每类资料放在哪、先看哪份汇总。问到作息、训练或学习安排时，它们先查真实记录和我写下的偏好与原则（40 多个主题页），再给建议。这些偏好与原则包括怎样算睡得好、什么状态适合训练，以及做选择时看重什么。Claude, Codex, ChatGPT and Hermes (an off-the-shelf assistant I chat with on Telegram) all start from the same guide file: where each kind of record lives and which summary to read first. When I ask about sleep, training or study plans, they check my real records and the 40-odd topic pages of preferences and principles I have written, such as what counts as good sleep for me, when I am ready to train and what I weigh when choosing, before they answer.
系统分三块：我的记录，存实际发生的事，也就是上面这些；我的想法和原则，是我确认过的偏好、经验和原则；给 AI 的说明，告诉它先读什么、能用哪些工具。下面的演示里，这三块分别叫数据核、认知核和 AIOS。The system has three parts: my records, what actually happened (the records above); my thinking and principles, the preferences, experience and principles I have approved; and the guide for AI, telling it what to read first and which tools it can use. In the demo below they are called the Data Core, the Cognitive Core and AIOS.
网页上能同时看到当天的睡眠、身体恢复和屏幕时间，以及本周的课程；早报用的也是这套计算。截图是 2026 年 10 月 7 日的真实界面，文字和数值已模糊。The dashboard shows the day's sleep, recovery and screen time alongside this week's classes, and the morning brief uses the same calculations. The screenshot is the real page on 7 October 2026, with text and values blurred.
看拆出去的 Health.AIO，有可试用的演示 →See Health.AIO, which has a demo you can try →
个人数字分身 第四版 概念演示
我的记录、判断与工具，在同一个问题下协作。
例如：先为重要任务留出连续时间。
再决定哪些经验留下。
Apple 健康
手机 App
Mac 屏幕
使用记录
手表 → 手机
→ Mac → 服务端
天气 API
导入的
个人 3D 模型
示意：列出记录可以从哪些地方来，不代表每一项都已经接入。
README 与执行边界
相关记录与认知
适合本次任务的技能
通过网页、Telegram、Claude / Codex 与我交互
整理过的内容与经历
已确认的个人取舍
认知之间的连接
内容来自第一版、第二版写在 Word 里的文字，之后一直在整理；新的认识由我确认后留下。
演示以安排一天为例，内容是示例，不读取我的真实数据。The demo plans a day with sample content and does not read my real data.
iPhone 采集 App 的签名过期，健康数据停了五天。页面照常显示旧数据，看上去一切正常。The iPhone collection app's signing expired and health data stopped for five days. The pages kept showing old data, so everything looked fine.
加了每天 9 点的数据检查，哪一路断了就发消息告诉我；早报也写明数据停在哪天，不再拿旧数当今天。I added a daily 9 am check that messages me when any source stops, and the morning brief now states the last day with data instead of passing old numbers off as today's.
自己做的聊天界面和让几个 AI 分工合作的框架越来越重，要维护的东西越来越多，很多和我真正想解决的问题无关。The chat interface and the framework for getting several AIs to split work, both built by me, kept growing, and much of the upkeep had little to do with the problem I actually wanted to solve.
2026 年 9 月起，聊天和定时推送交给现成的 Hermes 和 Telegram，手机 App 只管采集。我的记录、想法和给 AI 的说明文件仍然单独保存，以后换工具也不用跟着搬。From September 2026, chat and scheduled messages moved to Hermes and Telegram, and the phone app only collects data. My records, my thinking and the guide file for AI stay separate, so changing tools later does not mean moving them.
2026 年 10 月 8 日，我出了 25 道情境题，每道题事先写好必须答到的要点。判分交给另一个隔离的 AI 会话；回答里的睡眠时长这类数字，用系统自己的工具重新取一遍核对；“像不像我”的题由我自己打分。首轮通过 19 道。On 8 October 2026 I wrote 25 scenario questions, each with the points a good answer must cover. A separate, isolated AI session graded the answers; measurements such as sleep duration were re-fetched with the system's own tools and checked; questions about whether an answer sounded like me were scored by me. The first round passed 19.
没通过的题里，有一部分是 AI 该查资料时没查，直接按常识猜；最弱的是“猜我会怎么选”这类题，3 道只过了 1 道。后面要改的就是这几处：规定什么时候必须先查；引用要对得上原文，数字必须来自工具；缺数据时直接说缺。Some failures came from AI guessing from general knowledge when it should have looked things up. The weakest type was predicting how I would choose: 1 of 3 passed. The next changes target exactly this: when it must check first, quotes that match the source, numbers that come from tools, and saying plainly when data is missing.
还没用全套题重新测过，所以 19/25 只是首轮结果，不代表现在的准确率。The full set has not been rerun since, so 19/25 is a first-round result, not a measure of current accuracy.
HRV（心率变异性）常用来看身体恢复得怎样。我想让系统从第二天的 HRV 倒推出前一天哪件事做得不好，于是让 AI 比了四类模型，包括神经网络；数据是 2018 年以来 HRV 记录比较完整的约 2,300 天。HRV (heart rate variability) is a common sign of recovery. I wanted the system to work out from the next day's HRV which of the previous day's choices hurt it, so I had AI compare four kinds of models, including a neural network, on about 2,300 days since 2018 with reasonably complete HRV records.
单件事对第二天 HRV 的影响不到 1 毫秒，日常波动却有 ±5 毫秒，要几百个晚上才分得出来；神经网络也没有赢过简单的线性模型。唯一稳定的一条是多睡：多睡 1 小时，第二天 HRV 平均高 0.3–0.8 毫秒。我据此决定，网页不去猜“昨天哪件事让今天变差”。A single choice moves HRV by less than 1 ms while everyday swings are about ±5 ms, so it would take hundreds of nights to tell them apart, and the neural network did no better than a simple linear model. The one consistent signal was sleep: an extra hour went with 0.3–0.8 ms higher HRV the next day. So I decided the dashboard would not guess which of yesterday's choices made today worse.
多睡和 HRV 的关系只是观察数据里的关联，没做样本外验证。The sleep–HRV link is an association in observational data and has not been validated out of sample.
把自己的想法、原则和经历写进 Word。这批文字后来成了“我的想法和原则”这一块。I wrote my ideas, principles and experiences in Word. That text later became the “my thinking and principles” part.
重新整理了一遍内容和关联，但文字只能写我想怎样生活，看不出实际做到了没有。于是需要另一份记录，专门记“实际发生了什么”。I reorganised the content, but text could only describe how I wanted to live, not whether I did it. I needed a second record of what actually happened.
用网页整理身体和训练记录，画出变化趋势。可记录、计算和显示都挤在同一个页面里，每改一处都牵动别处。A web page organised body and training records and charted the trends. But recording, calculation and display all lived in one page, so every change touched everything else.
把我的想法和原则与我的记录分开存，加上给所有 AI 的同一份说明，再接上 iPhone 采集、网页和每天的早报。I kept my thinking and principles apart from my records, added one guide for every AI, then connected iPhone collection, the web dashboard and the daily brief.
2026.10：健康管理部分拆成开源 App Health.AIO；个人数字分身继续管其余的记录、想法和 AI 协作。Oct 2026: health management became the open-source app Health.AIO; the Personal Digital Twin keeps the other records, my thinking and the AI collaboration.
AI 提出的新理解先放进候选区，我确认后才算数；一条新记录也不能直接改掉我的长期看法。另外，偏好只能解释我为什么这样选，不能当成发生过的事。A preference can explain why I chose something, but it is not a record of what happened. New interpretations from AI wait in a candidate area until I approve them, and one new record cannot overwrite a long-held view.
原始记录只往后加，要更正也是另加一条修订，所以什么时候变过都查得到。Raw records only grow. A correction is added as a new revision, so every change can be traced.
图中的数量按 2026-10-07 统计；检索结果来自当时的索引测试，不代表现在的规模，也不代表 AI 回答的准确率。Counts are as of 7 Oct 2026; the retrieval results come from an index test at the time and show neither today's scale nor how accurate the AI's answers are.
示意：四个 AI 都从同一个入口进来，按编号只取这次任务要的几页；圈出来的就是这次取的页，蓝圈是正式内容，琥珀圈是还没确认的候选。我用 20 道固定测试题检查它能不能取到对的页（测的是检索索引，不是 AI 的回答），最近一次（2026.07）20/20。Illustration: all four assistants come in through one entry point and load, by ID, only the pages a task needs. The ringed pages are the ones read this time: blue rings are approved pages; the amber ring is a candidate I haven't approved. A fixed set of 20 test questions checks that the right pages come back (it tests the retrieval index, not the AI's answers); it passed 20/20 at its last run (Jul 2026).
示意：AI 只能提出候选（琥珀空心点）。我确认之后，它才写进正式内容；不确认，就一直留在候选区。Illustration: an AI can only propose a candidate (amber, hollow). It joins the approved pages after I approve it; until then it stays among the candidates.
据 2026-09-09 的架构图简化，补上了 9 月下旬做的网页端，其余按当时的情况画。认知和记录之间的编号关联只在样例上验证过，还不能说记录已经能自动对上认知。Simplified from the 9 Sep 2026 diagram, adding the web interface built in late September; everything else is drawn as it was then. ID links between beliefs and records have only been checked on samples; records cannot yet be matched to beliefs automatically.
从 1 开始：1 → 2 → 3，再按需读取 4 / 5。采集支线是 6 → 7 → 4 → 8；9 是候选确认支线。编号用于阅读，不代表所有模块依次执行。Start at 1: 1 → 2 → 3, then read 4 / 5 as needed. The collection branch is 6 → 7 → 4 → 8; 9 is the candidate-approval branch. Numbers guide reading, not a single execution sequence.
左右滑动，查看完整架构图Swipe sideways to explore the full diagram
上面四个问题的改法都由我决定，写成规格交给 AI 实现，做完再放回日常使用里检查。I decided how to handle each of the four problems above, wrote the specs for AI to build, then checked the result in daily use.
有些功能是从我自己的问题来的。比如我想知道玩游戏多了，会不会越来越难专注；手表测不了这种变化，我就让系统用娱乐时长和最长一段专注时长的 7 天平均，放在 90 天里对照着看。Some features came from my own questions. For example, I wanted to know whether more gaming was making it harder to focus. A watch cannot measure that, so I had the system plot 7-day averages of entertainment time and my longest focus block side by side over 90 days.
让新记录自动对上我写下的想法，出现反例时提醒我重新考虑，目前只在样例上试过。AI 查到资料后能不能用对，也还要靠情境题持续检查。Automatically matching new records to what I have written, and prompting me to reconsider when they disagree, has only been tried on sample data. Whether AI uses what it finds correctly still needs ongoing scenario tests.
Health.AIO 从个人数字分身拆出，训练后恢复是 AIO 里的一个具体场景。Health.AIO was split out of the Personal Digital Twin; post-workout recovery is one use case inside AIO.
项目之间的经验迁移Learning carried between projects
Ready to Train 状态评分与人体酸痛图Readiness rating and soreness body map AIO 晨间自评AIO morning check-in
Ready to Train 仍是独立的课程项目，部分设计已复用到 AIO；进一步的升级思路写在它的项目页里。Ready to Train remains an independent coursework project; parts of its design are already reused in AIO, and its project page sets out how to take them further.