What happens to your AI chats? Retention, training and how Nivana handles them

Where your words go

A chat message makes at least two stops. It goes to the app you are using, which may save it so your history syncs across devices. Then the app sends it to a model provider (Anthropic, OpenAI, Google and so on) to generate the answer. Each stop raises the same three questions: is it stored, for how long, and can it be used to train future models?

Zero data retention and no training

Nivana reaches models through Vercel AI Gateway, which offers two controls that matter here. With zeroDataRetention set, a request is routed "only through providers that have zero data retention agreements with Vercel AI Gateway".¹ With disallowPromptTraining set, requests are "only routed to providers that do not use your data for training".² Nivana sets both on every text request to a language model, including the small helper calls that classify and summarise. A provider without both agreements is simply never used for those calls.

What Nivana stores, and how

• Chats in the apps are saved to your account so they sync between the web, iPhone and Mac. They are encrypted at rest with AES-256-GCM, the authenticated mode NIST standardises in SP 800-38D.³ The key for each account is derived with HKDF, the standard key-derivation function in RFC 5869,⁴ so no two accounts share a key.
• Folders and settings are encrypted the same way.
• API requests are not stored. Nivana keeps routing metadata (which model answered, the task, token counts, timing and cost, never the content) to bill you and tune the router.
• Long-chat summaries are not written to storage. Each one is sealed to your account and travels back to your app with the answer. See why AI forgets long chats.
• Models never see your account. Providers receive a private pseudonym instead of your 1Profile identity.

Private chats

Start a chat as private and it is never saved to your history or synced. On the web it is forgotten as soon as you move on from it; on iPhone and Mac it lives only in the app's memory. The model still has to read your messages to answer them, and those text requests get the same zero-retention and no-training treatment as any other.

The anonymiser

For chats that mention real people, Nivana can replace personal details with stand-ins before a model ever sees them, then swap the real details back into the answer, so you read your colleague's name and the model only ever sees an alias. It is off by default and costs a few credits per message when it is on.

• What it catches: email addresses, phone numbers, card numbers (checked with the Luhn algorithm, so random digits are not mistaken for cards), US Social Security numbers, IBANs and IP addresses, plus people's names. Names are found with compromise, an open-source natural-language library,⁵ and a few rules about cue words such as "my colleague" or "Dear".
• What it leaves alone: well-known people, places and organisations. Asking about a head of state or a city is not personal data, and replacing it would ruin the answer.
• Consistency: the same person gets the same alias throughout your chats, so the model can still follow who is who.
• Limits: it is pattern-based, so it can miss an unusual name or an identifying detail that is not a name. It runs in the Nivana apps, not on API requests.

Turning models off

Some teams cannot send data to a particular lab, and some people simply prefer not to. In Settings you can turn off any model in the pool. The router then never chooses it, for your answers or for the helper steps behind them, which fall back to another allowed model.

Questions to ask of any AI app

• Are my prompts used to train models, by the app or by the providers it calls?
• Does each provider keep my prompts, and for how long?
• Is my saved history encrypted, and can I keep a chat out of it?
• Which services fall outside those promises (images, voice, search)?

Every assistant should answer these clearly. For Nivana, the details are in the privacy policy and the data and privacy docs.