SINGAPORE, 30 AUG 2026 — OpenAI has changed what a ChatGPT temporary chat does. A temporary chat can now read your saved memories, custom instructions and plugins, and you can convert one into a permanent chat afterwards. Neither was previously true.
What actually changed
Temporary chats now come in two forms, and the form is fixed when the chat is created.
A personalised temporary chat uses your existing memories, custom instructions and plugins to tailor its answers. It does not create new memories, and it stays out of your chat history unless you choose to save it. A non-personalised temporary chat behaves as before: no memory, no custom instructions, no plugins, no new memories.
Saving is the second change. Where a conversation turns out to be worth keeping, you can add it to your chat history, at which point it becomes a regular chat and follows your account-level personalisation settings from then on.
Personalisation cannot be changed mid-conversation. Whatever you chose at the start is what that chat is.
Reading and writing are now separate
Memory now has two directions that can be controlled separately: reading from it and writing to it.
The old temporary chat blocked both. Nothing came in from your history and nothing went out to it, which is why the feature was easy to reason about: it was a sealed room.
The new personalised mode blocks only the write direction. Your accumulated context flows in, so the model knows what it knows about you, while nothing new is recorded. That is genuinely useful, because a private question answered with the benefit of your stated preferences is a better answer. It is also a different privacy model from the one temporary used to mean.
OpenAI is not hiding this. The behaviour is documented and the choice is presented at creation. The risk is that a user who learned what temporary chat meant in 2024 will not re-read the documentation in 2026, and the name has not changed.
What saving actually does
The save option is worth thinking about before you use it.
A saved temporary chat stops being temporary in every sense. It joins your history, it becomes referenceable by future conversations, and it falls under your account-level personalisation settings, which means that if memory is enabled it can contribute to what the model knows about you going forward.
The decision to save is not merely filing. It moves a conversation you had deliberately outside your permanent record into that record, with all the consequences for retention and personalisation. That is a reasonable thing to offer and it deserves a moment's thought rather than a reflex.
Note the asymmetry: a temporary chat can become permanent, and a permanent one can only be deleted, not demoted back.
Why the feature was changed
The stated reason is user feedback: sometimes a temporary chat turns out to be worth keeping. It is an obvious complaint. Anyone who has lost a good conversation to the wrong mode has felt it.
The product logic underneath is also clear. A sealed mode is a mode in which the assistant performs worse, because it cannot use what it knows. Every session a user spends in it is a session that demonstrates the product at less than its capability, and there is a commercial incentive to narrow the circumstances in which that happens.
Offering personalisation as a choice is the honest way to resolve that tension, and it is what has been done here. The alternative — quietly enabling memory in temporary chats — would have been a different story.
The residual concern is about defaults and naming rather than about intent. A feature whose behaviour is chosen at creation depends entirely on the user understanding the choice at the moment it is offered, which is exactly when they are thinking about their question rather than about their settings.
Who this matters to
For most personal use the change is an improvement and the exposure is small.
It matters more for people using temporary chat as a compartment: a journalist working on something sensitive, a clinician thinking through a case, an employee asking about a situation at work, anyone whose reason for the mode was that the conversation should not join the rest.
For those uses the correct setting is the non-personalised mode, chosen deliberately at creation, and the correct habit is not to save. The personalised mode gives up the property that made the feature attractive to exactly this group.
Enterprise deployments have their own layer of controls, and an organisation with a policy on ChatGPT use should establish whether personalisation defaults are set at the tenant level or left to each user, because a default that individuals can change is not a control.
The wider pattern with these products
This is the second capability change in a month that widens what ChatGPT can reach, and the direction is consistent.
We covered the plugin that lets ChatGPT read your iMessages, and what full disk access means. That was about a new source of data flowing in. This is about an existing store of data flowing into a context that previously excluded it.
Neither is a scandal and both are the same movement: the assistant works better the more it knows, so the product keeps making more of what it knows available in more places. The response is not alarm or indifference, but a habit of noticing when features keep their names while changing their behaviour. Re-read what a setting does about as often as the product ships.
If you use temporary chat for a specific reason, check which mode you are getting before your next sensitive conversation rather than after it.
View author profile → · Editorial policy
