Your data

Your conversations, your questions, your notes and your uploads are not used to train AI models. This is not per-engine fine print — it covers every model on the shelf, wherever that model was built.

Enough for the product to be a product, and it is worth being plain about which parts those are:

• Your conversations, so they are there when you come back and so a thread can be continued.
• Artifacts and versions, so a deck you were editing is still the deck you were editing.
• Notes and uploads you chose to bring in.
• Memory, when it is on — and you can turn it off, or make Vresk forget all of it, from Your data.
• A message you have typed but not sent, kept on your device so a reload doesn’t lose it — it never leaves your device and we never see it. It goes when you send it, clear the box, delete your data, sign out, or when anyone else uses this browser. On a shared or family device, that last one means a draft can disappear before you come back to it — which is the trade we chose, because it also means the next person never finds your words.

Images are re-encoded on upload — the stored file is never the original bytes, and camera metadata, location included, does not survive the re-encode.

Saved conversations are also what makes the shelf work at all: an engine can only pick up your context because your context lives in your account rather than inside a model.

All of these are in your account’s privacy settings, and all of them take effect without asking anyone:

• See where your data goes. A disclosure list built live from the engine catalog and your own origin setting each time the page loads — it cannot drift from the engines actually on offer.
• Choose retention. By default a conversation with no new message for 720 days (about two years) is removed; one switch keeps your conversations until you delete them. No uploaded file is removed for its age, either way; a guest account’s files go with the account when it is removed, including by the guest cleanup once that is switched on. The switch does not touch deletion: anything you delete is deleted for real.
• Turn memory off. With it off, Vresk saves nothing about you between conversations and uses nothing it saved before — memory has its own page.
• Take a copy. One file: your conversations and messages, notes, memories, projects, settings, and a list of the files you uploaded — and the file states on its face what it leaves out.
• Delete all your data. Removes what you have made and keeps the account. You type a short phrase to confirm, and the server checks it too. If part of a delete cannot finish, it says so — running it again finishes the job.
• Delete your account. Everything above plus your settings, connected tools and the account itself, behind the same typed confirmation. A few records outside your account stay, and the privacy policy lists them. If we are ever legally required to preserve specific material, for example in connection with a child-safety report, we will keep only that material for as long as the law requires.

Open your privacy settings.

A safety check runs around model turns. It is not something you can switch off, and it is the one place Vresk deliberately sits between you and the model. Everywhere else the aim is the opposite — to get out of the model’s way, because the model is not the part we built.

If a message trips the self-harm check, the reply leads with a card of support resources — rendered first, above anything else in the reply. The card is deliberately calm: it asks you for nothing, tracks nothing, and its links carry no referrer.

When a check declines a message, the notice names what happened — our check held it back — rather than passing a verdict on you; the mechanics are on the troubleshooting page.