Deleting your AI chat history

The three different operations

1. Removing it from your view

The delete icon next to a conversation. Instant, visible, and the weakest of the three. It takes the conversation out of your sidebar and starts a deletion on the provider's side, which finishes over a period of days rather than immediately. Useful if the risk you are managing is someone opening your laptop. Not the same as the data ceasing to exist.

2. Turning off training

A switch in the data controls, usually phrased as improving the model for everyone. Turning it off means future conversations are not used to train the model.

Two things it does not do, and they trip people up constantly. It does not delete anything already stored. And it does not stop the conversations being retained. Retention, training and visibility are three separate settings, and flipping one is routinely mistaken for flipping all three.

3. Erasing the account

The most thorough option, because it targets the identity rather than a list of records, and it is irreversible. Export anything you want to keep first, because you will not get a second chance at it.

What deletion does not reach

This is not a claim that providers are dishonest. It is how any large system works, and the exceptions are written in the policies if you go looking.

• The window. Deletion is generally queued rather than immediate, and completes over days. In that period the data exists.
• Safety and abuse retention. Providers commonly keep a copy for a limited period to investigate misuse. It sits outside the normal deletion path by design.
• Backups. System backups are written on their own schedule and expire on their own schedule. A record deleted from production can persist in a backup until that backup rotates out.
• Legal holds. If a provider is subject to a preservation order, deletion can be suspended entirely, including for users with no connection to the case. This is not hypothetical and it is not something a user can see, opt out of, or be notified about.
• What you already sent. Anything a conversation went on to produce, such as a shared link you created, is a separate object with its own lifetime.

Deletion is a request to a system you cannot inspect, subject to exceptions you are not told about when they apply. That is the honest description, and it is why the timing of the decision matters more than the thoroughness of the cleanup.

Doing it properly, in order

The exact menu paths move with every release, so here is the shape rather than a screenshot that will be wrong in a month. In every mainstream assistant it lives under settings, in a section called data controls or privacy.

1. Export first, if you want any of it. Every major provider offers a data export. It takes a while to arrive and you cannot ask for it afterwards.
2. Turn off model training. Do this before deleting, so the window between the two is not spent contributing.
3. Shorten retention if the setting exists. Some products let you set conversations to expire automatically. Where it exists it is the most valuable switch on the page, because it works on everything future without you remembering.
4. Delete the conversations you care about, individually or all at once.
5. Check the connected surfaces. Mobile apps, browser extensions, and any workspace or third-party integration you connected. Each may hold its own copy or its own cache.
6. Then decide about the account itself. If the reason you are here is a specific conversation you regret, deleting the account is the version that also removes the identity it was attached to.

If you are in the EU or the UK

You have a right of erasure under the GDPR, and it is stronger than the in-app button: it is a legal obligation on the company rather than a feature it chose to offer. You can make the request in writing to the provider's data protection contact, and they generally have one month to respond.

Two caveats, said plainly. It is a request with a process and a wait, not a switch. And it does not undo anything that has already happened, which is the general problem with all remedies of this kind: they work after the fact, and the exposure you are worried about happened at the moment you typed.

The version that always works

Everything above is cleanup. It is worth doing, and it is worth being clear about what kind of thing it is.

A record that does not exist has no deletion window, no safety copy, no backup rotation and no legal hold. It cannot be produced under a request, exported by an administrator, or caught by a retention policy that changes next year. Not creating it is not a stronger version of deleting it; it is a different category of protection, and it is the only one that does not depend on someone else's process working correctly.

In practice that means preferring, for the questions that actually matter to you, an assistant that has no account to store them under and keeps history on your own device.

Where NOLO stands

NOLO has no sign-up. The first visit generates a random identifier such as nolo_a7b3_x9k2 and that is the login, so there is no profile holding your conversations and nothing on our side to submit a deletion request about.

Chat history is written to your browser's local storage on your own device. Clearing the browser is the whole deletion, it is immediate, and there is no queue and no exception, because there is no second copy on a server to reach. Anonymous mode goes one step further and does not write the conversation to your browser either, which is the mode for anything you would otherwise be deleting afterwards.

What NOLO's servers do hold, in full: a pseudonymous daily message count against that random identifier, kept 30 days so the daily allowance can be enforced, and rate-limit buckets keyed to a hashed IP address, kept 7 days. No chat content and no readable IP address. Both expire on their own without you asking.

And the limit, stated here as everywhere else: the text of your message does leave your device to reach a model. Fast and Plus run at Groq with zero data retention enabled on NOLO's account; Pro and image reading go through OpenRouter with zdr and data_collection: deny in the request, so the platform refuses to route to any provider that would keep it. That is a shorter path with fewer copies, not an invisible one.

Questions people ask next