What Edmund does with your note
Triage is the whole trick. You capture without thinking, and this is the part that does the thinking for you, in the background, before you next look.
Save a note and it goes quiet for a second or two. In that gap Edmund reads what you wrote and works out four things about it.
The four guesses
It writes a title, so a two-line ramble becomes something you can scan in a list later. It picks a type: a task, an idea, a bug, a chore, and so on. It picks the project it belongs to, using your project descriptions. And it adds a few tags.
Your original words are never overwritten by any of this. The title sits on top of what you actually said, and you can flip between the two on the note, or write your own title if Edmund’s isn’t to your taste.
How sure it is
Every note comes back with a confidence, shown as the little bars on the note. Above about seventy percent Edmund is reasonably happy. Below it, you get a Review pill, which is Edmund telling you plainly that it had a guess and wouldn’t bet on it.
That pill is doing real work. It’s what lets you skim an Inbox of twelve notes, confirm nine without reading them properly, and give your attention to the three that need it.
Which engine does the thinking
There are three, and Edmund picks the best one available to it without asking you.
On a recent iPhone or Mac it uses Apple’s on-device models. Free, private, works in a tunnel with no signal, and good at the job. If you’ve added your own Anthropic key it uses Claude instead, which is a bit sharper on long or messy notes. On older devices it falls back to a basic match, which reads the words rather than the meaning. That one is noticeably less clever, so it deliberately never claims high confidence, and you’ll see the Review pill more often.
The note itself tells you which one ran, under the title. If the cloud engine can’t be reached because you’re offline, Edmund quietly drops to the on-device one rather than sitting there failing.
It learns from you correcting it
When you change a triage guess, Edmund keeps a note of it and feeds similar corrections back in as examples next time. Move three notes about invoices into Admin and the fourth tends to land there on its own.
This stays on your device, and it only kicks in with a key set, since the on-device models don’t take examples the same way. Nothing about your corrections goes anywhere.
When it gets it wrong
It will, sometimes, and the design assumes so. That’s why nothing is filed or sent onward off the back of a guess: everything waits in the Inbox for you to confirm. Fixing a wrong guess is one tap on the note, and it makes the next one better.
If it’s getting a lot wrong, the answer is almost never in the AI settings. It’s in the project descriptions, which is the one place your input genuinely changes the results.
Next: confirming notes in the Inbox, or back to Help.