Security & privacy

It reads your email. That demands receipts.

A product like this earns trust with specifics, not badges. Here is exactly how your data is handled — and the hard limits that don't move.

Your data is not our productContent is processed by our AI providers, under their API terms, solely to serve you — drafts, briefs, risk checks. We never train our own models on it, never sell it, never use it for ads — and we're moving every provider onto no-training agreements as we scale. It works for you, not on you.

Personal email is excluded in codeNot a setting. Not a toggle. Personal conversations are structurally untouchable, and that rule cannot be configured away — by you or by us.

Credentials are encrypted at restWe never see your password — sign-in is Google OAuth only. Access tokens are encrypted (Fernet) before they touch a database, which itself encrypts at rest. All transport is TLS.

Tenant isolation at the database layerEvery row of your data carries your user id, enforced with row-level security. Your memory is yours alone — there is no cross-customer pooling of any kind.

Uncertainty

What happens when it isn't sure.

The rest of this page is about what we do with your data. This part is about what the AI does when it does not have an answer — and it is the part worth reading closely. Every claim check returns one of four verdicts.

SupportedYour own record backs the claim. No flag is raised and nothing interrupts you.

Cleared

Partially supportedSome of it checks out and some of it doesn't. You get both halves, separately, rather than an average of them.

Partial

ContradictedYour record says otherwise. A flag is raised with its sources attached, and autonomous sending freezes on that thread.

Frozen

UnverifiableNothing in your record speaks to this either way. It says exactly that — it does not round the gap up to “fine”.

Unknown

The fourth row is the one to look for in any product like this. A system that can only answer “fine” or “problem” will answer “fine” when it has nothing to go on — and “fine” from a system that could not check is indistinguishable from “fine” from a system that did.

The same rule runs through the product. A contact with too little history reads “not enough history”, never “healthy”. Mail no layer has read yet reads “not analysed”, never zero findings. Absence is stated, and it is never dressed up as an all-clear.

The specifics

Fair questions, straight answers.

What exactly do you store?

Work-email metadata and content needed to serve you, plus what the AI derives from it: facts, commitments, and relationship signals — each with its source and date. It lives in our cloud database (hosted in Japan) isolated to your account; the full list of processors that touch it is in our Privacy Policy. You can browse all of it in the Memory view; nothing is hidden.

What about information concerning my contacts?

Memories are derived from correspondence you already possess, stored privately in your account, and used only to assist you. Deleting your account deletes every memory referencing every contact, immediately and cascade-complete. If a contact asks you for removal, deleting their contact entry removes their derived memories — or write to us and we'll handle it.

Can I leave with my data — or wipe it?

Both, in one click each. Export downloads everything as JSON. Delete permanently removes your account and every derived row — memories, scores, commitments, drafts, logs — in a single cascading operation. No retention games, no 30-day limbo.

What stops the AI from sending something it shouldn't?

Four hard limits: it starts draft-only and autonomy is granted by you per category; every send has an undo window; any active agreement-risk flag freezes sending on that thread entirely; and sends respect your working hours. The first and third are enforced in code, not preferences.

Which providers can connect today?

Google Workspace / Gmail and Google Calendar are live. Microsoft 365 / Outlook is on the roadmap — we know Google-only is a constraint, and reducing single-provider dependency is planned engineering, not a maybe.

Compliance posture?

GDPR-aligned by construction: data minimization, full export, cascading deletion, EU-standard processing terms. SOC 2 is on the roadmap as we move upmarket; security review questionnaires are welcome — ask and we'll answer honestly, including about what isn't built yet.

Trust it with a look first.

Start in draft-only mode. Watch what it catches. Grant autonomy only when it has earned it.

See what it finds

One-click Google sign-in · no credit card.
Questions first? Talk to us — security questionnaires welcome.