How it works

Written to be checked, not believed.

Most tools tell you what they do. This page tells you how, precisely enough that you can catch us being wrong. If something here stops matching the product, that is a bug and we want to hear about it.

The one idea

Search works when you remember you saved something. The hard case is the other one: you read the right article eight months ago and you have no idea it exists any more. So while you write, the notebook looks at what you are working on and puts the relevant part of your own past material beside you — without being asked, and without you remembering anything.

What happens when you save something

The note is split into passages of at most 1,500 characters, on section boundaries where it has them. Passages, not whole notes, are the unit throughout — the point is to hand you the paragraph, not a document to re-read.
Each passage is turned into 384 numbers that stand for its meaning, by a model that runs in your browser. The passage text is not sent anywhere to do this.
Those numbers are kept so the comparison below can happen in memory, without a round trip to a server.

What happens while you write

When you pause, the notebook takes what you have been writing and turns it into the same kind of numbers.
If that is more than 92% similar to the last thing it looked at, it does nothing — you have not changed subject, so the answer has not changed either.
Otherwise it compares against every passage you have saved and keeps the closest ones, at most one per note, favouring variety over five near-identical passages. Recency is a mild tiebreak, not a ranking.
Anything scoring below 0.45 is discarded. If nothing survives, you get nothing. At most five cards appear.

Where 0.45 comes from

It was measured, not chosen to sound good. On a real 280-note notebook, genuine matches for a question scored 0.47 to 0.55, and unrelated everyday sentences topped out around 0.37. The floor sits above the noise and below the signal. It is deliberately closer to the noise than is comfortable: we would rather miss something than fill the sidebar to look busy.

This is the number most worth arguing with, and the one we expect to move. If the rail is silent when it should not be, that is the dial.

Passages that are yours to open

A passage can be exactly relevant and still be the wrong thing to put in front of someone. If you write about a hard week, the closest match in your notebook may be the worst week you ever recorded. The matching is working perfectly and the result is awful.

So a passage that reads as your ownaccount of loss, illness, an ending or real distress is not hidden and not ranked lower. The card still shows where it came from and when. The words wait behind “Show it”. Two things must both be true — it is written in the first person and it is about one of those subjects — because an article about grief is not the same as your own entry about it.

The honest limit: this is a word list, not a model. It will hold back ordinary notes sometimes, and it will miss things. We accept the first because the cost is one click, and we would rather be wrong in that direction. It is not a safety feature and makes no clinical judgement.

What runs on your device, and what does not

The decision about what deserves to interrupt you happens on your machine. So does turning your notes into searchable meaning, transcribing your voice, reading your imported files, and reading text out of images in the desktop app. None of that sends your notes anywhere.

Your notes are stored in your account, which is why your phone and laptop show the same notebook. Two things use a hosted AI model: the assistant, when you ask it a question, and topic tags, which are off unless you switch them on. Cloud re-ranking and summaries are opt-in and labelled where you turn them on. The full list is in the privacy policy.

What it will not do

It will not invent anything. Every card is something you saved, with a way back to where it came from.
It will not fill the space. No results is a valid answer and the common one on a small notebook.
It will not ask you to keep a habit. No review schedule, no daily queue, no streak.
It will not make you file anything. Things from one source gather into one note on their own.
It will not charge you to save. Capture and search are free and unlimited, permanently.

How to check us

Take your notes and leave. Export gives you plain files that open in anything. If that ever becomes hard, we have broken the promise.
Watch the network while you type. The rail does its work without asking a server anything. Only the assistant makes a request.
Test the floor. Write about something you have never saved. The sidebar should stay empty. If it finds something anyway, tell us — that is the failure we care most about.

What we are still unsure about

Whether 0.45 is right for a notebook much larger or smaller than the one it was measured on. Whether the passages held back are the right ones. And the real question underneath all of it: whether the rail can become trustworthy enough that its appearing is itself a signal worth looking at, rather than one more thing to ignore. That is what this beta is for, and it is why the cards ask whether they helped.