From the founder. Written without a name on it, because a name would make you weigh the argument by who is making it — and the argument here is the only thing that should have to stand up.
There is a drawer. You probably have one too.
Mine has the notebooks in it — the good ones with the elastic band, the cheap ones from the stationery shop near the office, two from a job I left years ago. Somewhere in that drawer is a number I worked out carefully one evening, the name of the person who offered to introduce me to someone, and about four ideas that were better than anything I've had since.
I know they're in there. I have never once gone and found them.
That's the whole problem. Not writing. After writing.
So I bought an e-ink tablet, the way you do. And I want to be fair to it: it is a genuinely lovely object. The pen feels like a pen. The screen doesn't glow at you at eleven at night. Writing on it is better than writing on paper in several small ways I didn't expect.
Then I filled it, and the drawer moved inside the device.
Same problem, better housing. I could scroll instead of flip, which is faster in the way that running in a slightly larger circle is faster. Search worked inside one notebook if I had asked it to convert that notebook to text, which I hadn't, because you have to decide that in advance and I never know in advance which page is going to matter.
The Android one can run anything, which turns out to be a way of saying it is a slow tablet wearing a notebook's clothes. I put a note app on it. Then another. Then I was choosing note apps instead of writing notes.
The third had the best paper feel of the three and the most thoughtful software of the three, and it still could not tell me what I had concluded about a thing across two notebooks and four months, because none of them can. I checked. Someone wrote a teardown of all three in 2026 and put it better than I can:
“None of them offer cross-notebook full-text search. None produce searchable PDFs. None export Markdown. All three companion apps offer: nothing. No REST API. No outbound webhooks.”
Three devices. One problem. Nobody had touched it.
This is the part that turned an annoyance into a project.
I went looking for other people complaining about it, expecting to find a handful. I found twenty-one separate threads spanning 2018 to 2026, asking the same question in almost the same words. Eight years. The tablets got thinner, the pens got better, the marketing got more serene — and every year someone posted the same question and got the same non-answer.
One person, describing why they'd stopped bothering, wrote the sentence that ended up shaping the entire design: every note arrives with a question attached — which notebook does this belong in — and that question is the expensive one.
They were right. It's why I stop writing things down. Not the writing. The filing.
So gemnotes doesn't ask. You write. Anywhere, in any order, badly. The sorting happens afterwards, on a computer, where there's room to see everything at once and where being wrong is cheap because nothing was moved.
I assumed I'd find a good technical reason. What I found was an incentive problem.
The market leader spends about 4.9% of revenue on R&D, and a chunk of that goes into internal business systems, not the product. That is a hardware company's number. A hardware company sells you the next device — and making your old notes more valuable gives you one less reason to buy it. Nobody has to be cynical for this to happen. The incentive just quietly points the other way, every year, for a decade.
And neutrality is worse than unprofitable for them: it's unthinkable. reMarkable will never read Boox files properly. Boox will never read Supernote's. Meanwhile the person with three tablets — me, and it turns out a lot of other people — needs exactly the layer that no manufacturer is allowed to build.
That's the gap. Not a clever feature. A structural hole that four companies each have a reason to leave open.
I started this wanting to build a device. Of course I did — the object is the fun part, and I had opinions about margins and pen weight I still think are right.
Then I looked at the numbers. In FY2025 the best company in this category lost $35.6 million on $423.7 million of revenue, cut around 40% of its staff and replaced its CEO. The first cause they named was memory prices — DRAM rose 90–95% quarter-on-quarter in early 2026.
If the company with three million devices in the wild and the strongest brand in the category is losing money making these, my building one first is not courage. It's a bet dressed as one.
So the device is designed, rendered, specified — and parked, behind four written conditions that have to be true before a single panel gets bought. Writing them down in advance was the only way to stop myself re-arguing it on a good day.
The software comes first. It's the part that actually solves the drawer.
When a competitor launched something similar into this community, the top-voted comment on the announcement was a refusal — on privacy grounds. Not a feature complaint. A refusal.
I understood that immediately, because my notebooks contain the version of my thinking that exists before I've made it presentable. That's the whole value of writing by hand: it's the only place you're honest. Uploading it to someone's cloud to make it searchable is a trade where you hand over the thing that made it worth keeping.
So the models run on your machine. It costs us a 10.9 GB download, it costs us people whose laptops are too small, and it means we cannot use the biggest models. Those are real costs and I'd pay them again.
Writing, syncing, recognition and search work. I use it. Ink latency on a Boox is 5 ms, handwriting comes back at 87.5% of words exactly right, and on a real archive run, three pages that had never been findable by text came back on the first three queries I tried. That was the day this stopped being a research project.
The diagram rebuild — the screen this product is really for — does not work yet. The first attempt at the stroke classifier scored zero. It's written on the status page in those words, next to everything else that isn't finished, because I'd rather you find that here than after installing it.
The drawer is still full. But for the first time in eight years, something is reading it.
Every figure in this piece — the 21 threads, the 4.9%, the $35.6 million, the 5 ms, the 87.5% — is on the status page with the date it was measured or the source it came from. If a number here can't be backed there, the number is wrong and we want to know.