Your email, kept safe
Ariadot reads some of the most personal things you own, so you deserve to know exactly how we handle it. Seven promises, in plain words, and one honest limit we won't hide from you.
i.
You sign in through Google or Apple, the same way you sign in to anything else. Your password never leaves them. We can't see it, type it, or store it.
ii.
Google only lets us see what's already in your inbox. We have no way to send mail, reply to anyone, or move a thing to the bin. Not by accident, not on purpose. That limit is set by Google's read-only permission, not just promised by us.
Got a work inbox and a personal one? Connect both, and Ariadot reads them together into a single morning note, without ever crossing the read-only line on either.
iii.
Before we ask an AI to help write your note, we change every real name, phone number, address and dollar amount into a believable fake one. The AI reasons about a stranger's details, never yours.
We use realistic stand-ins, not robotic placeholders like «person_3», for two reasons: models reason better over natural text, and the fakes are drawn from a synthetic set that's uncorrelated with anything real. So even if a model's output leaked, it would describe a coherent stranger, not you. The swap is also stable: once "Mary" becomes "Jackson," she stays Jackson in every future note, which is what lets Aria reason about your week without ever seeing the real cast.
We swap the real ones back at our end, just before the note reaches you, using a map that's sealed with a key unique to you and never sent to any model. What we don't touch are dates and times, because a note that reasons about the wrong day is worse than useless.
iv.
We mostly use Claude, made by Anthropic, with a couple of smaller models for the quick sorting and screening. Each one only ever sees the swapped, fake version, and none of them keep it or learn from it.
Your email is also untrusted by default. Every message is checked for hidden instructions (the "prompt injection" trick where an email tries to talk to the AI) before any reasoning, so a message can never tell Ariadot what to do.
v.
A guard reads every message first and drops anything that looks like a password, one-time code, API key, or recovery phrase. It's never stored and never sent to a model. Ariadot is about your loops and your dates, never your secrets, so we built it to refuse them on purpose, not just decline to ask.
vi.
Once Ariadot has read a message and written your note, the original is deleted within a day. Most go within minutes; a daily sweep makes sure nothing older than that remains. Your email still lives safely in Gmail, where it always has, so there's no reason for us to keep a copy.
What we do keep is the useful part: your loops, your handbook, and a quiet profile of your routines, all encrypted at rest. And you can see every bit of it in the app, including an activity log of what Ariadot decided about each email it read. Nothing is hidden in a back room you can't look into.
vii.
Settings → Delete account. There's a 7-day window in case you change your mind, then it's gone for good: every loop, note, and handbook entry, the profile Aria built, and your Google connection switched off at Google's end. No "we'll keep it just in case."
What it will and won't keep
Ariadot keeps loops and a handbook. It is not a password manager, and the line between the two is drawn in three tiers, checked before a single thing is written down.
No friction. This is the handbook.
How it behaves. Saved when you save it. Surfaced when you ask, or quietly alongside a related deadline. The everyday details that are a pain to dig up, right there when you need them.
Stored once you say yes, per category. An identifier labels an account; on its own it unlocks nothing.
How it behaves. The first time one comes up, Aria asks: "This is an identifier, not a password. Store it? Yes / No / Always allow this category." Never read out in the morning note, only fetched when you go looking.
No override. If possession alone unlocks something, it belongs in a password manager.
How it behaves. Caught and refused before storage. Aria won't keep it, and it never reaches a model. The same check runs as a fast pattern match and again as a model pass, so it's defence in depth, not a single filter.
Common worries
Warm answers, and we don't dodge the hard one.
You're fine. Nothing important lives on the phone in the first place. Sign in on a new one and your briefs, loops, and handbook are all right there waiting, because they live with us, not on the device. A lost phone loses nothing.
No, and there's no door for us to do it through. There's no internal tool that opens your inbox, so no one at Ariadot, engineers included, sits down and reads your mail. The only thing that ever looks is software, making one quick pass and moving on.
A lot is working in your favour here. The only copy that ever leaves our edge for an AI is the swapped, fake-name version, so the model is reasoning about a stranger. The raw email is gone within a day. Your access tokens are sealed with a key unique to you. Everything is encrypted in transit and at rest. That's several independent locks, not one.
The honest part: to write your note, our software does read your email in readable form, and we hold the keys to do that, so we're not a vault that only you can open. We'd rather tell you that plainly than pretend otherwise, and then build every layer above so it doesn't come down to one.
You'd get plenty of warning and a clean exit. We'd give you 30 days' notice by email, let you take whatever you'd like to keep, then wipe everything. No quiet shutdown, no wondering where your data went.
Just tell Aria what was off, and the next note is better for it, that's how it gets to know you. And on the rare day something breaks and we can't write one, we say so plainly. No invented notes, no pretending the day is handled when it isn't.
That's where we're headed. For now Ariadot is built around one person, the one carrying the household's mental load. Shared handbooks for partners are coming soon, so the household details can live in one place, and fuller household sharing follows once we can do it really well, not just quickly.
The exact model list, retention mechanics, sub-processors, and your legal rights live in the full privacy policy. The FAQ covers the rest.