How it works
Last updated 4 September 2026
The front page is the short version. This is the long one, for anyone who wants to know what is actually happening before they trust it with fifteen years of notes.
Getting the notes off your devices
Every device SageArchive supports has its own way of handing over a note, and each one is taught to the app separately. What they all share is how the app treats them: it asks for a page of your history, writes those notes down along with a bookmark saying how far it got, and only then asks for the next page.
That sounds like a detail. It is the reason you can close the lid halfway through syncing four years of recordings, open it tomorrow, and have it pick up exactly where it stopped rather than starting over or quietly skipping the notes it was in the middle of.
A folder of audio files works too, and needs no account at all. Drag it in. If the same recording arrives twice — once from the device and once from an old backup — the app notices it is the same audio and keeps one.
Writing it down
Transcription runs on your Mac using a well-regarded open speech model, accelerated by the same chip that runs everything else on the machine. There is no upload, no queue behind other customers, and no per-minute charge.
We measured it under load rather than in a demo: an hour of continuous work, fifty-one hours of audio processed, about fifty times faster than real time — and it did not slow down as the machine got hot. That is the number that matters for a backlog. Several hundred hours of old recordings is one night's work.
Long recordings are cut into pieces before processing, and the cuts are made at the quietest moment near each boundary, so no sentence is ever sliced in half. The timings are then shifted back onto the real clock, which is why clicking a result three hours into a meeting lands where you expect.
About who said what
Speech models that run on your own machine cannot tell voices apart — that is a limitation of the technology, not of this app, and no local option at any price does it today. If a particular meeting really needs "who said what", you can opt that one recording into a cloud service that does it, and the app will tell you before it sends anything.
Reading handwriting
Handwritten pages are read twice over. First by macOS's own text recognition, which is good at telling us where on the page each line sits. Then, optionally, by a stronger model that is better at telling us what the line says. The two readings are lined up against each other so that the better text inherits the right position.
Roughly two lines in three come out with an exact position. The app draws a highlight only on those, and marks the rest with a small dot meaning "found and read, but we could not place it precisely." We would rather show you a gap than a confident box in the wrong place.
Finding things
Four searches run at the same time on every query:
- The exact words you typed.
- Near misses — so a typo, or a name transcribed slightly wrong, still finds the note.
- The meaning — so "the renewal conversation" finds a note that never used the word renewal.
- Structured things — IP addresses, network ranges, hardware addresses, port numbers, hostnames and dates.
That last one exists because ordinary search is terrible at it. Search for 10.11.99.1 in most tools and you get every note containing the numbers 10, 11, 99 and 1. SageArchive treats it as one thing. It will find a machine inside a range you searched for, and the range that contains a machine you searched for. It knows that 28 Jan, 1/28 and 2026-01-28 are the same day, and when a date is genuinely ambiguous it keeps both readings rather than picking one and being silently wrong.
The four sets of results are then combined by position rather than by score. That is a dull-sounding choice with a real consequence: combining raw scores from different kinds of search quietly re-weights itself as your archive grows, so a tool tuned on a thousand notes gets worse over a year without anyone changing a line of code. Combining by position does not drift.
Answering questions
When you ask a question rather than search for a phrase, the app finds the most relevant passages from across your whole archive, hands only those passages to a language model, and asks it to answer using them.
The important part is how the citations work. The model is never allowed to write a timestamp or a page number. It refers to passages by an internal label, and the app turns that label back into a real position afterwards. If the model refers to something that was not in the passages it was given, that reference is thrown away before you ever see it.
This is the difference between a citation and a plausible-looking citation, and it is the single thing we were least willing to compromise on. A tool that invents a timestamp once is a tool you can never trust again.
By default the answering model runs on your Mac too, using the model Apple ships with macOS. It is a small model and it shows: it cites correctly but tends to quote its sources rather than synthesise across them. If you want a stronger answer you can turn on a cloud model for that purpose — and only that purpose — and the app will show you exactly which passages would be sent.
What "nothing leaves your Mac" actually means
There is no SageArchive account, no SageArchive server, and nothing for us to be subpoenaed for. We could not read your notes if we wanted to, because they never come near us.
Cloud services are available for the two jobs where they genuinely help — transcription with speaker labels, and stronger answers — and both are off until you turn them on, one device and one purpose at a time. The app is built so that a piece of your data can only be sent after that permission has been granted; a version of the app that leaked by accident would not build in the first place.
Your archive lives in a folder on your Mac that only your account can open. The audio and page images are kept as ordinary files outside the database, so backing up does not mean copying forty gigabytes every time one note changes.
What it is not
SageArchive is not where you write. There is no rich-text editor and there is not going to be one. It reads what you captured somewhere else, and it is very good at finding it again. If you love Obsidian, or Notion, or a paper notebook, keep them — the intention is to sit underneath whatever you already use, not to replace it.
It also does not build knowledge graphs, extract entities, tag your notes for you, or write summaries of everything the moment it arrives. Those were considered and deliberately left out. Each one is a way for a tool to spend your battery being confidently wrong about your own material.
Requirements
- macOS on Apple silicon.
- Disk space for your own audio, plus a few gigabytes for the speech model, downloaded once.
- An internet connection only for downloading the app, the model, and whatever your devices need to sync. After that it works with the network off.