TentanoteMost people who care about their notes have a graveyard. An Evernote account last opened years ago. A Roam graph from the winter everyone started one. A folder of Markdown files with a system in it that made sense for about three weeks. Each one began as a promise, to put everything in one place and find it again, and each one ended the same way.
The tool was a promise, and the sorting was homework. Everything went in, but it only came back out if someone had tagged it, linked it, or put it in the right notebook, and that someone was you. You did it for a while. Then a busy month came, the inbox filled up, and the pile stopped being worth searching because you no longer trusted it to be complete. Once you stop trusting a pile you stop adding to it, and at that point it is dead.
None of this was a failure of design or effort. Software could not do the one thing that mattered. It could store a receipt, but it could not tell you the receipt belonged to the kitchen renovation. That took a person, so a person had to be the librarian.
What changed
That constraint is gone. Software can now read a receipt, an email thread, a calendar invitation or a page you saved at midnight, and say what it is and what it probably belongs with. It does this well enough that asking a person to file by hand is now a choice rather than a necessity.
And filing was never quite the job. Most of what lands on a Mac is noise: the newsletter you meant to read, the fourth confirmation for the same flight, the tab you opened once and never looked at again. Somewhere in the same pile is the signal: the builder's quote, the three people you keep meeting, the page you have come back to four times. Some people enjoy keeping a library. Almost nobody enjoys deciding, one item at a time, which of the two a thing is. What software can now do is tell them apart from what you do rather than what you tag: who was there, what you came back to. Keep the signal. Let the noise fade into the background, not into the bin, so it is still there the day a search needs it.
So the obvious pitch writes itself: the first version of this idea where you do not have to be the librarian. The machine files, and you get on with your life.
That is half right, and the other half is where these tools go wrong next.
What didn't change
A machine that files will sometimes file things wrong, in ways only you can see. Here is the shape of it. You have three calls with the same building firm in a fortnight, and the software proposes the firm as a project. It has read the data correctly: there really are three calls with one company. But the company is not the project. The project is the kitchen, and the builder is one of a dozen things attached to it, alongside a tile shop, a loan and an argument about a sink. Or an old contact card of yours, from a job you left, links five unrelated calls because each of them mentions your old email address.
Neither is a failure to understand the data. Each is a correct reading that reaches a wrong conclusion about a life. And the obvious remedy, having the machine propose and letting you approve, feels responsible and is a trap. It moves the homework instead of removing it. Nobody files, and nobody works through a queue of the system's guesses either.
A review queue is the Evernote inbox with better reasons attached.
A badge with a count on it is a chore with a number on it. People stop opening it for the same reason they stopped tagging: it asks for attention on the system's schedule rather than theirs. And a map of your life that you do not trust gets abandoned faster than an empty one.
Making a wrong guess cheap
The way out is not to make the machine right every time. It is to make it cheap when it is wrong, and to stop asking you to vouch for it. These are the rules Tentanote is being built to:
- Nothing waits for approval. What the system concludes stays marked as its own, with its reason, for as long as it exists. It is never a pending item. No badge, no count, no inbox of proposals. Someone who never reviews anything should still end up better off than someone who never filed anything.
- Guesses change what surfaces, not where things live. What the system infers decides what shows up together when you look for something, so a bad guess costs a slightly worse result, not a mess. The moves that are hard to undo, like making a project, merging two things or moving one, are left to you.
- Corrections happen where you run into them. Nobody checks a queue, but everyone reacts to a wrong result in front of them. Dragging something out of a place it does not belong fixes that result and teaches the system in the same gesture. It costs nothing you were not already doing.
- Questions come only when you are already organising. Some things only you know. Nothing in a calendar says house hunt. So the time to ask is the moment you make a project called House Hunt yourself: eleven viewings with four agents look like they belong here. One answer files eleven things, and it comes when you already meant to tidy up.
- No means no, for good. A system that brings back what it was told to drop is nagging.
There is one exception. The first time you open it, it shows you what you already have, connected, with a reason on every line, and all of it stands unless something in it is wrong. It is a reveal, not a review.
Two numbers to keep honest
If these rules hold, two things can be measured. How often something the system surfaced gets undone, which says whether it is guessing. And how many questions it asks you in a week, which should trend towards zero. A rising first number means the machine is wrong too often. A rising second means it is handing its uncertainty back to you, and that is the homework coming back in a new form.
The honest version of the promise is smaller than the pitch, and better:
The machine organises quietly, you correct in passing, and nobody is asked to do either as a job.