How to keep an AI chat private and temporary
Temporary Chat hides a conversation from your history and deletes it within seven days. Here is what that covers, and what it cannot.
A letter arrives from your insurer about a claim, written in language you have to read several times. You photograph it, open a chat, and ask a model to tell you what it actually says. Then you hesitate, because the page in front of you now contains your policy number, a diagnosis and a sum of money, and it is about to become a saved conversation.
The hesitation is reasonable, but it usually bundles three separate worries into one. Will this sit in my history where anyone who picks up my laptop can read it. Will the company that made the model keep it. Will it end up training something. Only the first is fully in your hands, and Temporary Chat, which shipped on web and mobile in version 1.0.3, answers it cleanly. How far it reaches into the other two is worth knowing before you lean on it.
What a temporary chat switches off
Temporary Chat is a mode you switch on for one conversation at a time. Turn it on and the chat never appears in your sidebar, your thread search, your folders or your archive. It is excluded from every list surface in the app, which is the part people actually want.
Underneath, more is switched off than the hiding. Saved memories are not read into a temporary chat, and nothing from it can be written back into memory. No title is generated for it, so it never gets a name that summarizes what you asked. Any file you attach goes to a separate storage area that is never promoted into your permanent uploads, and its bytes are handed to the model on the turn you send them rather than parked as a stored file on the provider's side.
What temporary does not mean is that nothing was written down. A copy of the conversation lives on our servers for up to seven days so that abuse can be detected and acted on, and then an hourly job deletes the thread and its messages outright. The seven days run from the moment the chat starts, not from your last message. The app says this to you before you type a word, in small print under the composer: "For safety, we keep a copy for up to 7 days to detect and prevent abuse, then delete it." Our privacy policy states the same seven days in its retention section, which is where to check it: product copy drifts, and a policy is what we are held to.
Turning it on, on web and on your phone
On the web app the control sits in the sidebar next to the New Chat button: a dashed circle that reads "Turn on temporary chat" when you hover it. Press it and the address bar stays on a plain temporary URL with no conversation id in it, while a small "Temporary chat" marker rides above the composer so you can still tell what mode you are in ten messages later. On iOS and Android the same dashed circle sits at the top right of the new chat screen, and a "Temporary" pill appears in the header once the conversation starts. Both platforms label the feature Beta.
The mode deliberately does not stick. Starting a normal New Chat clears it. The conversation's id is held only in memory, so reloading the page drops you on a fresh empty temporary screen rather than back into the old chat. That is a good failsafe if you forget to switch back, and it is also the trade: there is no route back to a temporary conversation you walked away from, even during the seven days.
The mode leaves metering entirely alone: a temporary message counts against your allowance like any other.
The half you do not control
Now the part that no toggle in our app can change. Your message still travels to whichever model you chose, and what that company does with it is governed by their terms, not ours.
MultiChats does not use your conversations, prompts or uploaded content to train AI models. That is our own commitment, written into the privacy policy, and it is a claim about us and nobody else. Be suspicious of any AI product that offers one flat sentence about training and stops there. Sixty models from eighteen different companies cannot share a single privacy posture.
So we put the difference on the model itself. Open a model's info card in the picker and, on the six models that carry a caveat, you get a line about data handling underneath the specs. Grok's is not softened:
xAI may use your conversations to train Grok models.
The other four read differently. Open-weight models routed through OpenRouter carry the notice "Routed via OpenRouter to third-party providers that may briefly retain requests; they do not train on your data." Twenty models in the catalog are additionally pinned on every request to a fixed list of six hosts: DeepInfra, Fireworks AI, Google Cloud Vertex AI, Amazon Bedrock, Groq and Microsoft Azure. That pinning decides who actually receives your prompt. Whoever developed the model, one of those six is the company serving it.
The practical instruction is short. Read the info card for the model you have selected before the insurer's letter goes anywhere near it. Temporary Chat does not change a word of what that card says.
Your profile travels with every message
There is one leak worth knowing about, and our own screen is part of why it catches people. Open a temporary chat and the app tells you it "won't be used to personalize your experience." That is accurate about your saved memories, which are a paid feature: they are not read into a temporary chat and nothing from it is written back to them. It does not cover the Personal preferences box in your account settings, the one whose own help text says your preferences will apply to all conversations. That paragraph, and the traits you picked next to it, go into the system prompt on every request we send, temporary ones included. There is no exception for temporary mode anywhere in that path, so the empty-state line is doing more work than it can carry.
So if the box names your employer, or the condition on that claim form, it travels out with the anonymous-feeling temporary chat too. Clear it, or keep it to things you are content to send with every message you will ever write here.
One detail runs the other way. If you signed in with Google, the name on that account is never forwarded to a model provider. Only a name you typed yourself into profile settings is used, and the settings screen tells you so before you save it.
Deleting a conversation, and how far it reaches
Delete an ordinary conversation and it disappears from your history immediately, and it is removed from our systems within thirty days, from backups within a further fourteen. Deleting your whole account erases conversations, files, generated images, memories, folders and preferences within the same thirty days. Two narrow exceptions survive that, and they differ. Where there is a payment dispute or suspected fraud on the account, we keep a usage record: timestamps, the model used, token counts and cost, with no conversation content in it. Where the case involves illegal content or a serious breach of the Terms, the conversations and files concerned are kept as well, so they can go to the competent authorities. The retention section of the policy sets out both, and is worth reading directly. Erasing any conversation, temporary ones included, leaves an hourly tally with no content in it (cost, number of replies, how many you deleted), kept up to 13 months for billing and refunds.
What deletion cannot do is reach backwards into the company that answered you. Your message left our servers the moment you pressed send. Removing your copy asks them for nothing, and whether they hold a transient log of that request, and for how long, is set by their terms. That is precisely what the notice on the model card is pointing at. The delete button is housekeeping on your own record, and it stops there.
When to skip temporary mode
Temporary Chat is the wrong tool whenever you want the conversation to still exist tomorrow. You cannot search it, you cannot return to it after a reload, it will never be titled, and branching from it produces another temporary chat with a fresh seven-day clock rather than something that lands in your archive. If the insurance claim turns into a month of correspondence you need to track, a normal chat with memories is what you want.
It is also the wrong lever for the worry people most often reach for it with. If what keeps you up is what a specific model provider might do with your text, hiding the chat from your own sidebar does nothing about it. The choice that moves that is the model you pick, on the info card.
So before you send the insurer's letter, do three things in this order. Read the info card for the model in the picker. Check what your Personal preferences box is still carrying. Then, if you want the chat kept out of your history, switch on Temporary Chat. That order matters, because the toggle is a strong answer to the question of what sits in your history and a weak one to the question of who reads your text, and the first two steps are the ones that touch the second question at all.