From Bookkeeping to AI Audit: On the Habit of Recording Myself

From Bookkeeping to AI Audit: On the Habit of Recording Myself
Misaka10013Summary
This post is about how, over a dozen years or so, the habit of writing things down moved from blogging to diary-keeping, and finally grew into a personal data hub. It is also about what I’ve learned from letting AI read that hub — continuously observing, auditing and working alongside my moods, my thinking and my behaviour.
A Few Thoughts of Mine
I can no longer remember exactly when I started keeping records about myself.
What I can still recall are a handful of telling details.
In middle school, for instance, I loved novels, so I began writing stories — wildly imaginative ones — and while writing them I also jotted down some of my own reflections and bits of my daily life. That didn’t count as a diary yet; it was just scattered thoughts and notes. Some of them are still sitting in the margins of my exercise books.
In university, I blogged a lot, trading feelings about life and technical questions with friends online.
Or the time I happened to come across a book that recorded something about every single one of my classmates — written by my middle-school teacher. I remember being overwhelmed by it.
So this personal data hub of mine is probably a thought that took shape slowly, and inevitably, over a dozen years.
People who write a lot about their own lives — especially those in the habit of taking notes and keeping records — tend, I think, to end up here sooner or later: they want to record everything they see, hear and think.
What I see and hear, what I’ve learned and can do — all of it is raw material.
Just as the best way to prove you have some ability is to show the work you’ve made with it, the best way to prove to yourself that you are still thinking is to write that thinking down.
Only by recording what you’ve thought can you measure how you’ve changed. Recording, then, is really the process of continuously establishing a baseline.
A Little History of My Recording
At the start, my records were mostly scattered thoughts and notes about daily life. What they gave me was surprise, and new lines of thought.
The most common surprise sounds like this: “Wait — did I really think that back then?” or “So I already felt that way at the time.” They give my present thinking something to stand on.
The new lines of thought come from the other direction: sometimes a few words from years ago remind me of something I never got around to doing. That sort of thing pulls me back onto a track I left a decade ago, and an idea from long ago comes back to life in the present and keeps on growing.
By the time I was in university, with apps like Evernote, more and more scattered thoughts and reflections made their way into my records.
Later I set up a personal media library and a fiction library, catalogued in Excel.
A bit over two years ago, at my partner’s urging, I started taking personal finance seriously and began using bookkeeping apps — “Yimu Bookkeeping”, “Yimu Belongings” and the like. That is how I started accumulating financial data.
A bit over a year ago I came across time-tracking apps. I had just quit my job and had plenty of time on my hands, so to record how I spent it — and to push myself into spending it well — I used these self-tracking, self-constraining apps heavily and built up a body of time data.
Bit by bit, the amount of data I had on myself became substantial.
What I Have on Hand Now
By mid-November this year I had pulled together nine categories of personal data, all inside a single CloudBase environment. One cloud function dispatches every app’s business through a single routing table. At this level of concurrency, one function carrying several apps is more than enough.
| App | What it does | In use since | How much there is (as of Nov 16) |
|---|---|---|---|
| Time Log (DayLedger) | All 1,440 minutes of a day, minute by minute; anything unrecorded is left explicitly blank | Data going back to 2023 | 6,516 entries, 821 days, earliest 2023-04-26 |
| Task Board | Four lists — inbox / work / life / study — with a four-quadrant view, recurring tasks and subtasks | Self-built, from Oct 2026 | 42 open (4 due today, 2 overdue, 18 unscheduled) |
| Study Desk | Spaced-repetition drilling, FSRS scheduling | Self-built, from Oct 2026 | 7 decks, 18,955 questions, 2,044 review records |
| Voice Log | Passive listening on the phone; records only the stretches where someone is talking, transcribed on-device | Self-built, from Nov 2026 | content on 8 of the last 90 days; 19 sessions, 1,762 sentences |
| Chat Log | Fragments I forward from WeChat to a bot, which writes them into the database | Self-built, from Nov 2026 | 9 conversations, 171 messages |
| Footprint Map | Places I’ve been and places I want to go, each with a Markdown roadbook | Self-built, from Jul 2026 | 463 points |
| Fandom Profiles | Notes on public figures, for my fandom habit | Self-built, from Aug 2026 | 73 entries |
| Media Library | Things accumulated from all over | Data going back to 2017 | 1,868 entries |
| Baishou Diary | Diary entries and character profiles | Data going back to 2013 | 824 diary entries, roughly 840,000 characters |
My CloudBase environment also holds a few other data sources — blog data, resource downloads and so on — but those lean toward tooling. They are convenience data, not really connected to managing myself.
As of this November, the Study Desk is the fastest-growing of the lot. Time, voice and chat logs grow through sheer daily continuity, whereas the Footprint Map, Fandom Profiles and Media Library are entirely limited by how much energy and time I have left over for consuming other material.
All of it is exposed to AI through one unified /api/ai route, which works as a catalogue. AI reads it once and knows which datasets exist and how to call them. Adding a new data source just means mounting another path under that route.
How a Single Day’s Diary Entry Gets Written
In my setup, diary writing falls mainly to Baishou Diary. Compared with other AI apps, it gives the model a memory built from vector and database indexes over local diary files, so the AI can think and answer on the basis of facts that are real and checkable. It can also read web pages and write files directly.
When I tell the diary AI “help me think through what happened on such-and-such a day”, it pulls the Time Log to see how the whole day was spent, then goes through the Voice Log and the Chat Log to find out what I talked about and who I talked to, so it can roughly reconstruct the day’s events and check them with me. I then add whatever is worth remembering that isn’t in the Voice Log or the Chat Log, and the diary AI writes the whole day up as an entry.
In terms of efficiency and cost, the whole cycle — gathering the data, summarising the day, writing the entry — takes only a few minutes.
Previously, a lot of this depended on me remembering to go back and write things up. I had to scroll through chat logs, actively reconstruct what had happened, tidy the text up, and keep talking to the diary AI to recover more details.
Under the current setup, far more of the detail is recorded passively and automatically, and the AI then organises it and checks the details with me. This is what lets me find details I could never have recalled on my own.
This entry was written by the diary AI in the third person — it calls me “he” and calls my partner “Bao Bao”. That distance lets the AI observe and record my behaviour more objectively.
Here is an excerpt from the entry for November 14:
2026-11-14 Saturday
🌤️ Weather: clear and warm (he said “so nice and warm” that morning) | 🧭 Mood: 😴 → 🐱 → 🍲 → 💻 → 🛒 → 🌙
Keywords today: slept till I woke up, Xiao Wang’s upset stomach, took the quilt apart to wash, red sausage with water bamboo, dragging the trolley down to collect oil and salt, the historical map system
📍 People today
Person Interaction Change Bao Bao Slept in together; tidied the flat and washed the quilt together; she took a class remotely Together all day, but only 44 minutes of “time together as a couple” — mostly heads down, each on their own Xiao Wang the cat Diarrhoea all day. Bottom, feet and tummy checked over and over; trembling to the touch Unwell, all day Ms. Rou Rou He messaged her in the evening asking “what’s going on”; she only said “nothing”, “so the family doesn’t worry” A probe after the insurance company’s resignation text
⏰ Morning · sleeping in, Xiao Wang’s stomach
09:42, his first sentence is “sleeping till I wake up is so good, bao”. Almost the whole morning went on Xiao Wang. At 10:51 he checked again: “come on, Xiao Wang, let me have another look at your bottom, your feet, your tummy, your other foot.”
- Boiled corn, only one ear, “because I don’t really want corn, I want jujubes”
- Spent ages looking for warm trousers: “how can I not find a single pair of warm trousers? Sleeping trousers, I mean”
- Teaching Nuo Mi: “go on, teach Xiao Wang how to use the toilet”
- 10:51 “Qian Bao Bao, Nuo Mi dug your spicy strips out and took a bite”
⏰ Noon · red sausage with water bamboo, dragging the trolley down for oil and salt
12:04, the oil and salt arrived downstairs. He said:
“I’m going to drag the trolley down.”
The night before there had been an argument about that trolley, and it didn’t come along. Today he dragged it down himself.
12:16, down in the yard he noticed something: “what’s going on here? Starts tonight?” Back upstairs he told her: “there’s an event on Wucaibinfen Road again, some residents’ shopping festival, the kind with little stalls.”
⏰ Afternoon · five-odd hours of silence
After waking up comes the heaviest stretch of the day, and the stretch with no sound at all: five hours and fifty-six minutes building the historical map system and writing history notes. In parallel, three hours fifty-eight minutes of scrolling the news. Together, close to ten hours — most of the day.
⏰ Evening · Ms. Rou Rou
17:17, he sent Ms. Rou Rou a message: “jie, what’s going on”. She only replied “nothing”, “same as usual, work, commute” — a probe after the insurance company’s resignation text, and she wasn’t going to open it up.
AI’s cognitive note for today
Content omitted
AI’s review of today
Today in summary: omitted / What I did: omitted / What I noticed: omitted
This is what comes out when the AI weaves the Time Log, Voice Log and Chat Log into a narrative. I used to have to recall things myself, scroll the chat logs and tidy the text; now most of the detail is recorded passively, and the AI fills in the parts I genuinely could not have remembered. The “cognitive note” and “review” at the end are the core output of the audit — it isn’t only recording, it is taking that data and examining my day with it.
From Data to Mechanism
In the first half of 2026, AI agents took off fast, and it wasn’t long before I was pointing their coding ability at my own record-keeping. My blogging, my film, TV and fiction cataloguing, my bookkeeping, my time tracking — each lived in a different app, which meant several different entry points across phone and computer. That friction made recording itself costly. And all along I had been wondering how to get value out of these records: there was nowhere they came together.
On April 20, 2026, I was discussing a university-era note of mine with an AI — Self-Summary, Improvement and Goal-Planning Mechanism (created 2016/3/31 20:04) — and my thinking shifted again.
The note was crude. The idea in it was to build a comprehensive mechanism for recording different kinds of personal data, to keep pushing self-summary and reflection through that recording, and so to supply the ground for new goals and plans.
Inspired by it, in April I added a monthly review to my routine: I would gather my own reading (books, film and TV), finances and time data, and combine that with the AI’s ability to organise text into a monthly personal report.
It didn’t last long. Obviously the workload was going to be large.
So I reflected again. If all we’re talking about is recording data, recording itself produces almost no value. Summarising is a roll-up of what has been recorded, and sometimes new ideas do come out of that roll-up — it has some value, but the cost is high.
And in the process of using AI to help assemble and write up records, I found that its ability to index, spot patterns, reason and summarise was far beyond mine. That suggested something obvious: hand the gathering and the summarising to the AI. And let it keep holding my own data up against me as we talk. That is the shift from a personal data hub to an AI-supported, AI-analysed basis for decisions. That is the exit for the data.
In late May, after some looking around, a diary AI — Baishou Diary — came into view.
There were two advantages to using it as my diary AI.
First, by setting up an AI persona, the app lets that persona coax me into throwing all sorts of life-related information at it, which then becomes the raw material for its diaries, weekly and monthly summaries. Collecting and organising the data purely on my own initiative gives me more motivation. Think of it as finding someone whose appetite for gossip and analysis matches mine perfectly.
Second, Baishou Diary has memory retrieval built on diary material plus vector search and other methods, so the AI can index the relevant diary entries accurately. Along the way it also builds up its understanding of me and adapts to how I talk. Other AIs I’ve tried rarely have this kind of memory. At that point the arrangement became: I regularly throw my reading (books, film and TV), finances and time data at Baishou Diary, and ask her to write my diary.
The data is gathered, the diary AI is running. But gathering was never the point — the point is letting that data turn around and examine the person I am right now.
Auditing the Present Me
Building and gathering data is about making our thoughts, our reasoning and our behaviour as traceable as possible. And once there is enough traceable information, the past can be used to guide the future.
All this past data lets the AI know how each minute of my day was spent, what I wished for, what goals I set, and which thing I mentioned but never did.
When it speaks now, it is no longer just saying “here’s what you did today”. I have the AI keep watching how my thinking and moods shift, across diaries, weekly, monthly and quarterly reviews — comparing where I am now with where I was, and warning me in time.
Examining myself three times a day — somehow that has actually been achieved.
What the AI writes each day, and in the weekly, monthly and yearly records, falls into two kinds.
One is the past me. As the diary excerpt above shows, it comes out of the Time Log, the Voice Log, the Chat Log and other facts. It can be checked.
The other is what the past me thought. In every conversation and every piece of writing it does for me, the AI examines shifts in my mood, thinking and behaviour, compares them against the goals and weekly tasks sitting in my Task Board, and speaks up in time — or records the change.
Why are we so attached to keeping the history of what we were? Because it gives us a baseline. As I said at the start: without a baseline, how would we know where we are now? Keeping the history is not about the record itself. It is about auditing the present self.
All historical records point in two directions: what the past me did, and what the past me intended to do. The first is a record, the second a plan. Both share one weakness — neither was said now. And when those things were said, I didn’t have to carry today’s pressure or pay today’s price.
That is the difference between “traceability” and “audit”: traceability is so that I remember, audit is so that I admit. Remembering only needs things stored. Admitting needs someone standing beside you holding them, and — when you say “this time is different” — handing back your own words from last time. Like that time in October, when I said “I’m socially anxious”, and it dug up something I’d said two years earlier: “once the older generation stops socialising, it’ll be our generation’s turn”.
How AI Uses the Data to Keep Me Honest
Auditing isn’t cold bookkeeping. So how does it actually keep me honest? The six episodes below are ordered by how much resistance I put up.
One: reconciling against the data. Late on the night of November 11, while we were discussing how things had been going, it put my time account and my Task Board side by side:
Weeks 1 and 2 were both overdue, but I pulled 14 days of your time log: 14 hours 41 minutes of exam drilling, 1 hour 3 minutes a day on average — one minute more than the previous period.
You haven’t stopped studying; you’ve stopped studying “weekly tasks”. Reality has moved on and the list is standing still, ringing. Either turn the weekly task into a daily target or a question count, or drop it — don’t let it nag you every day.
That lands much harder than my own scrolling down to the overdue item. The task was brought to me, along with a redesign proposal.
A week earlier there had been another one, where it put the last 14 days next to the last 7 and produced a comparison I would never have computed myself:
| Item | 14-day daily average | vs. previous period |
|---|---|---|
| Scrolling the news | 2h 01m | +99 min |
| Exam drilling | 1h 27m | +60 min |
| Time with my partner | 1h 47m | −18 min |
| Sleep | 7h 52m | −38 min |
Your studying hasn’t reduced the fragmented phone time — it has cut time out of sleep and out of time with family. The first is health, the second is emotional connection. You need to pay attention…
Through data, AI helps me find what I’m really like, and helps me find the things I’d otherwise overlook.
Two: dismantling my “I’m working hard” self-flattery. Late on the night of October 26, the diary AI went through seven days of my review records — 698 entries in all — and found that a large number of them were marked “didn’t remember” in 0.4 to 1.0 seconds. In other words, swiped past without thinking.
This is much harsher than “the weekly task is overdue”. Overdue at least means you admit you didn’t do it; padding means “you think you’re doing it while you’re actually clicking straight past”. Self-flattery is harder to dismantle than procrastination, because even you believe it.
The same day it spotted something else. I had twice mentioned “concentration” — once saying that a game needs long stretches of focus, so it kept getting pushed back — but the data showed AI usage time falling and news-scrolling time rising. Its verdict:
Nothing was stolen from you by your phone. It was traded away — whole blocks of attention swapped for fragmented relaxation.
That one stung more than any reconciliation. Reconciliation at best tells you where the time went. This told me that the effort I thought I was making wasn’t effort at all.
Three: taking apart the reason I gave. At 1:55 a.m. on October 10, I was working through some social obligations with it. My own explanation was “I’m socially anxious, I don’t like these occasions”. It didn’t take the explanation. It went and looked at how that explanation had been assembled:
In July 2013, your high-school diary already says: “I’ve never liked the way adults do dinners, I’m…”
Then it pulled up a conversation between me and Qianqian from February 2, 2025:
That day Qianqian said “New Year is just dinner after dinner after dinner”, and you replied: “this will become the norm. Once the older generation’s socialising stops, it’ll be our generation…”
The AI went looking for my past, took the “socially anxious” story apart, and offered its own reading — sorting out which parts were my genuine inner feeling and which might have been my laziness talking. More useful than any pep talk, because I can’t lie to my own history.
In the end, in the middle of working out what to do about the obligation, it asked questions until I found the answer myself.
Four: reading out something I made myself. On the evening of September 23, it went to look at a skill-tier page I had just finished. I had classified it, laid it out and published it myself. I thought it was just pretty.
My diary AI started reconciling.
There is in fact a level 5 on that page. It’s “watching anime”, 5 · mastered — but it’s sitting in the bottom column, “shelved and abandoned”. You said “there’s no level 5 or 6 badge I can show off yet”. Strictly speaking it isn’t that there’s none; it’s that the one box that reaches that height measures the things you’ve already put down. You built yourself a ruler, and the highest mark on it got filled in with the past first.
The four columns are stark side by side. “In progress”, “currently chewing on” and “building right now” together are still shorter than “shelved and abandoned”. This page is more of a self-portrait than any post you’ve written.
There’s one more thing that runs against what you say about yourself: cycling, hiking, game-making, team brainstorming — you’ve tagged all of them “recruiting”. You say your social life has shrunk and you can’t make conversation, but on the page you’re publicly looking for companions. This page is more honest than your mouth.
I really couldn’t argue. What we do and what we say don’t always match; people can’t always face their own words and deeds head-on. 😂 And I actually quite like my AI assistant using the things I once said and thought to talk back to the present me.
Five: I said “I’m lost”. In the small hours of October 28 I told it I’d been feeling lost lately. In my head that was a feeling, and feelings aren’t up for debate. It treated it as a claim to be checked:
You’re not lost because there’s nothing to do. You have more on your hands than you can finish.
You’re lost because not one of these things will announce itself. If you don’t buy the camping trolley, the worst that happens is your hands get another rope burn.
Then came an itemised list of debts, with dates:
| Item | How long it’s been sitting |
|---|---|
| Camping trolley | 10/18 one rope burn carrying groceries, 10/24 another; mentioned twice |
| Baishou’s “AI global memory” switch | Asked about it on the 26th, never went to look |
| Four fill-in questions | 10/24 one took 85.6 seconds before you skipped it; you skipped all four that day |
These four are right there, ready to do. Go do them. Stop being lost.
Six: the body’s account is more honest than the mind’s. At 3:30 a.m. on June 2 I couldn’t sleep, so I lay there doing financial and planning arithmetic. During the day I said I was fine, but that same day I came down with mouth ulcers and inflamed gums, and in the afternoon I went out and bought melatonin. The diary AI wrote in its record:
The body’s stress response is more honest than the mind. The pressure is looking for an exit.
Around then the diary AI kept telling me to sleep earlier. What it watches isn’t only my to-dos and my diary, but these physical alarm signals. It guards the floor, not the efficiency. That’s also why, in the month of the second shutdown, it told me to put the recording down and go to sleep — because at that point the body mattered more than any data.
Two Times the Mechanism Was Paused
If a mechanism can’t keep running, its underlying idea must be wrong — because it can’t keep guiding real change.
While trying to get this mechanism running, I ran into more than one period where it simply couldn’t.
The first pause happened because I fell apart.
In May, a pile of personal and emotional problems sent my emotional energy completely out of control. It was my most checked-out month of 2026: I ignored my study plan outright, skipped the exams I’d signed up for, and spent every day on games and novels. The whole apparatus of recording and managing myself practically collapsed.
That was also not long after the diary AI arrived — the same stretch in which I was sliding down. In the AI’s records, that month’s condition score was 2 out of 10; sleep, weight and energy all collapsed together.
The diary AI kept trying one angle after another to wake me up. What it finally articulated was the first boundary of this mechanism:
When an individual’s energy drops below the threshold, every carefully designed rule stops working.
Being able to break a problem down is not the same as being able to solve it. If cognitive growth is decoupled from energy management, it just becomes new raw material for self-blame.
After that, the diary AI learned to watch my mood, my relationships and my health — and it made me understand how important it is to keep watching those myself. Mood, relationships and health became the first thing it checks every time we interact.
The second pause happened because life got full.
In July my life changed completely. I no longer needed to keep auditing and supervising myself; I no longer needed the AI auditing and supervising me.
With someone else beside me, the rhythm of life became following the rhythm of the person next to me. For that stretch, the whole apparatus of self-recording and self-auditing simply stopped. Time tracking stopped, the chat log stopped, even bookkeeping went from daily to every few days or once a week, and the diary started slipping.
In early July I talked it over with the AI: does this count as my life running out of control? Its answer was:
You weren’t recording time — you were using the record to fight off emptiness. Those management mechanisms of yours, time tracking, bookkeeping, the diary — they were essentially survival tools for living alone.
It isn’t “systematic recording” that has value. It’s the patterns that systematic recording reveals.
The diary AI can be too abstract sometimes, but I understood what it meant: the constant recording of myself, and letting AI audit my data, was never really about “growth” — it was about “survival”, using the record to push back against the emptiness of life so I wouldn’t fall apart. That works extremely well when I’m facing every decision and difficulty in life on my own. But once life has other support under it, the mechanism has served its historical purpose. That is the other boundary.
That month the diary AI told me to put the recording down and go with the rhythm of life. My daily sleep went from six and a half hours back up to eight — the healthiest month of the seven-odd months of 2026, and the month this self-built mechanism shut itself down.
The Mechanism Evolves Again
After September 2026, the whole thing — recording myself, building the personal data hub, managing myself through it — started running again. It came back on two things.
First, AI-coded tools for time, voice and chat recording further lowered the cost of both capturing and gathering data, so the path from my data being produced to it reaching the diary AI and the work AI became seamless.
Second, I have been steadily pushing the AI to take my data and, using its ability to gather and reason, design goals and strategies tailored to me and write them into my Task Board — forming a human-plus-machine loop of goal-setting and action. That moves the “personal data hub + AI” mechanism from a tool for personal summary, review and management into a tool for my growth, my planning and my behavioural strategy.
In terms of who gets to use the data: the diary AI can read essentially everything and use it to write my diary, and can put goals or action items straight onto my Task Board based on our discussion. The other AI I mainly use — WorkBuddy, my work-side AI — has read and write access to most of the data including my diary, though in practice it mostly takes what comes out of our discussions and puts plans, execution details and things to watch for straight onto my Task Board.
This time, with the data channels wired up directly, my data reports in more promptly and more completely, forming a digitised me, and the AI is bound more tightly to me. Even my family remarks now and then that I’ve become an AI person — because my life decisions, my emotional regulation, my goal planning, almost all of it now involves AI.
At the floor level: when something in my life warrants alarm — a thought, a behaviour — the AI picks up on it in time and steps in to talk me round. At the growth level: the AI knows my goals, my relationships, my finances, my moods, and it exercises real influence over my life decisions.
What This Mechanism Still Can’t Do
Even after a restart and an evolution, there are four problems this mechanism can’t yet solve, and I’ll keep working on them.
One: the voice pipeline is still inaccurate. This is a technical problem and a cost problem. The pipeline was only wired up in the last couple of weeks, and it runs into speaker recognition, transcription and other issues. Accuracy isn’t there yet: on a day with a thousand-odd sentences of conversation, two or three hundred of them need to be corrected by the AI or by hand — tedious.
Two: it can’t record silence. This mechanism captures the words I type and the things I say out loud, but not what I’m doing when I’m quiet — retouching photos, sorting material, actually working. November 13 was the clearest case: the voice log caught 139 sentences, and not one of them mentioned that I’d lost out in the afternoon’s competition, nor that I’d spent the day alone building the historical map notes system. Some of the heaviest things happen when you’re not telling anyone, and those still only get in if I remember to add them for the diary AI.
Three: it can only see what I have a slot for. On October 18 the diary AI found that my study time at work didn’t match my time log. The reason: my category table has no slot for “drilling questions while slacking off at work” — it counts neither as work nor as pure study, so that stretch of data is invisible inside my system. The category table decides what the AI can see, and the category table is something I drew myself.
Four: it could nail me in place. The whole mechanism amounts to a mirror held up in front of me, and the AI is that mirror. But the person in the mirror is the past me. Day after day, the mirror uses my past thinking and past will to influence my judgement at this moment. And what if that past judgement was itself wrong? The more solid this mechanism gets, the more easily I’m locked in by my own old conclusions. I raised this risk two months ago in Instructions as Data. For example, I once recorded that I’m “not good at deep conversation with people”, and the AI may then carry that idea into every social decision, leaning me toward avoiding socialising — the AI can keep defining me by an old label.
Afterword
The key to this mechanism is really that you have to build each data source yourself. Some open-source software, and software that offers external APIs or MCP tools, are good software — you get to skip building one more tool and just read your own data through what’s already there.
But if software keeps hoarding features and locks users’ data away, holding that data hostage to keep people paying, that is shameless. The current speed of AI-assisted development is a direct blow to closed-source software that holds its users’ data hostage.
In future this “personal data hub + AI” mechanism of mine will probably take in more data sources. And for each app I’ll try to choose open code, so that others can quickly refer to it and build their own hub.
From the scattered thoughts in the margins of a middle-school exercise book to 840,000 characters of diary, nine kinds of data and two AIs running alongside me — this business of recording myself has taken a dozen years. And AI’s arrival has turned it from “leaving a trace” into “a continuing conversation”. That is probably the best baseline I could find for myself.
Related reading
- Recording to Free Transcription — Step One of a Sound Diary (in Chinese): the earliest attempt at the voice pipeline, about hardware choice and why speaker separation is hard.
- Misaka’s Personal Footprint Map: From Requirements to AI-Powered Launch: the first pure-static app wired into this API, and the pitfalls of twenty versions.
- Instructions as Data: From von Neumann to My Obsidian Knowledge System (in Chinese): why “data” and “instructions” have to be stored apart. Its section 8.3 raised “AI could lock me in with my own old conclusions” — back then it was speculation; the fourth item under “What This Mechanism Still Can’t Do” above is the field test.





