You open a free ChatGPT account, type a few questions about a job offer and a breakup, and assume the chat ends when you close the tab. As of June 2026, it does not: OpenAI's new "Dreaming" memory now builds a running profile of you in the background, and it is rolling out to free accounts for the first time. You should know what it captures before you decide whether to leave it on.
Most coverage frames this as a feature upgrade: ChatGPT remembers more, so it helps more. That is half the story. The half that matters for your privacy is that the profile is now inferred and self-revising, not a tidy list of things you asked it to save. You are no longer auditing what you told it. You are auditing what it concluded about you, and quietly rewrote, on its own.
What is actually new in June 2026
OpenAI began rolling out the "Dreaming" memory system on June 4, 2026, announced as "Dreaming: Better memory for a more helpful ChatGPT." It launched first for Plus and Pro users in the US, with free and Go users added over the following weeks. The reason free accounts get it now: OpenAI says it cut the compute needed to run dreaming for free users by roughly 5x, which is what finally made a free-tier rollout practical.
Dreaming runs as a background process that synthesizes memory across many conversations, so context that surfaces naturally in a chat gets captured without you asking it to remember anything. It also updates old memories as time passes. OpenAI's own example: "You're going to Singapore in July" becomes "You went to Singapore in July 2026" after the trip ends. Recall on OpenAI's internal eval rose from 41.5% in 2024 to 82.8% in 2026.
The profile is inferred, not dictated
Here is the part that changes the privacy math. Older ChatGPT memory was mostly an explicit list: you said "remember that I'm vegetarian," it saved a line, you could read it back. Dreaming flips that. The system now decides, on its own, what is worth keeping and how to phrase it. A February 2026 research paper cited in coverage of the rollout found that 96 percent of memories are created unilaterally by the system rather than at the user's explicit request.
So the profile is not a transcript of what you typed. It is a set of conclusions: your likely job, your tone, your relationship status, the topics you keep circling back to, your apparent location and time zone. Inference is where the risk lives. A model can be confidently wrong about you, and it will still act on that guess in every future chat until you find and correct it.
You are no longer reviewing what you told ChatGPT. You are reviewing what it decided about you, and rewrote, while you were not looking.
What the inferred profile actually contains
OpenAI does not publish a fixed schema, but the memory page and the rollout examples make the categories clear enough. Dreaming captures stable facts you mention (your role, the tools you use, projects you keep returning to), preferences it pieces together (how you like answers formatted, what tone you respond to), and time-bound context it then ages out, like the Singapore trip in OpenAI's own example. It also tracks location and time zone signals so recommendations match where it thinks you are.
None of that is inherently sinister. The catch is that the same machinery picks up things you never meant to register as facts about you: a question asked on behalf of a friend, a hypothetical you were stress-testing, a one-off mood. Because dreaming synthesizes across conversations rather than copying single lines, those signals can harden into a stated trait. That is the gap between a memory list you curate and a profile a model writes.
There is a memory page where you can see what ChatGPT may use to personalize replies, add or correct facts, and dismiss entries you do not want. Useful, but limited. Tech Times reported that the new architecture provides only a limited audit trail, so you can see the current state of a memory but not always when or why it was revised. That gap matters most precisely when the inference is wrong.
Memory and model training are two different switches
This is where most people get burned: turning off memory does not stop your chats from training OpenAI's models, and opting out of training does not stop the memory profile. They are separate controls, in separate settings menus, with separate defaults. You have to handle both.
On training: for personal Free, Plus, and Pro accounts, the setting that lets OpenAI use your conversations to improve its models is on by default. It lives in Settings, Data Controls, as "Improve the model for everyone." Toggle it off and new conversations stop being used for training. Opting out does not delete past chats or pull back data already used in completed training runs, and a temporary chat is the option that skips both training and memory for a single session.
Why does the distinction matter so much here? Because the two systems compound. The memory profile makes ChatGPT keep a model of you across sessions, and the training default lets your raw conversations help improve the product for everyone. People assume that flipping the memory toggle off handles the privacy question. It does not touch the training pipeline at all, and the training opt-out menu is one most free users never open. If you only do one of the two, you have solved half the exposure and may believe you solved all of it.
| What you want to stop | Where the switch is | What it does |
|---|---|---|
| The inferred profile forming | Settings > Personalization | Turn off Reference saved memories, which also turns off Reference chat history. No new memories form. |
| Your chats training the model | Settings > Data Controls | Turn off Improve the model for everyone. New chats stop feeding training; past data is unaffected. |
| Both, for one session | New chat > Temporary chat | That session is excluded from memory and from training, and is not saved to history. |
How to turn ChatGPT memory off, exactly
Open Settings, then Personalization. ChatGPT memory runs as two linked features: "Reference saved memories" (the synthesized profile) and "Reference chat history" (pulling context from past conversations). Turn off Reference saved memories and Reference chat history switches off with it. From that point, no new memories form.
Switching memory off does not erase what is already stored. To clear the existing profile, open Settings > Personalization > Manage memories and delete entries there, or tell ChatGPT to forget specific things.
Two snags catch people here. First, the controls sync across your devices, so flipping the toggle off on the web also disables it on mobile, which is what you want but worth knowing if you share an account. Second, if you only want personalization paused for a single conversation, you do not need to dig through settings at all: start a temporary chat, and that session is left out of memory entirely while your standing settings stay where they are.
Why deleting a chat does not delete the memory
This trips up almost everyone. Saved memories are stored separately from your chat history. Delete the conversation where you mentioned something sensitive, and the memory ChatGPT derived from it can still survive and surface in future chats. The chat is gone; the conclusion it drew is not.
To actually remove something, you delete it in both places: the saved memory under Manage memories, and the original chat (plus any files or connected apps that hold the same information). With Dreaming inferring and rephrasing memories on its own, one fact you shared can end up worded differently in the profile than how you typed it, which is exactly why scanning the memory page beats trusting that a deleted chat cleaned up after itself.
Should you turn it off?
Use ChatGPT for low-stakes, general work? The memory is a convenience. Leave it on, audit the page occasionally, and opt out of training. But if you discuss anything you would not want inferred, stored, and potentially used to improve a vendor's model (health, finances, legal questions, work under NDA, anything about other people), the cautious default flips: turn memory off, turn training off, and use temporary chats for the sensitive sessions. On a free account, the secure choice is now opt-out, not opt-in.
A practical middle path works for most people. Leave memory on for the everyday stuff where personalization genuinely saves time, open the memory page once a month to delete entries that are wrong or too personal, and keep the training toggle off so your conversations are not feeding model improvements regardless. Reach for a temporary chat the moment a session turns sensitive, and remember that a temporary chat is excluded from both memory and training but is also not saved to your history, so copy out anything you want to keep before you close it.
A profile you own instead of one a vendor infers
The deeper issue is not one toggle. It is that the profile lives inside OpenAI, is built by an inference engine you do not control, and on a default free account can feed model training. MemX takes the opposite stance: it is an external memory layer for ChatGPT, Claude, and Gemini that is private by architecture, with per-user isolation and encryption at rest, and your memory is not used to train anyone's model. You decide what gets remembered and what gets removed, in plain entries you can read, rather than auditing a guess a model made about you. When you delete something, it is gone, not split across a chat store and a separate memory store you have to chase down twice.
That is the difference Dreaming makes concrete. One model holds a self-built picture of you that updates without telling you why. The other holds the facts you chose to give it, where deleting an entry actually deletes it. If the June 2026 rollout made you check your settings, that instinct is the right one to build on.
Frequently asked questions
01Is ChatGPT memory on by default for free accounts now?
Yes. OpenAI began rolling out its Dreaming memory system to free and Go users over the weeks following June 4, 2026, after a roughly 5x compute cut. On a default account, memory and model training are both on unless you switch them off.
02How do I turn off ChatGPT memory?
Open Settings, then Personalization. Turn off Reference saved memories, which also turns off Reference chat history. No new memories form after that. To delete what is already stored, use Manage memories in the same menu.
03Does turning off memory stop my chats from training the model?
No. They are separate controls. To stop training, go to Settings, then Data Controls, and turn off Improve the model for everyone, which is on by default for Free, Plus, and Pro personal accounts. Memory settings do not affect it.
04If I delete a conversation, is the memory gone too?
No. Saved memories are stored separately from chat history, so a memory can survive after you delete the chat it came from. To fully remove it, delete the entry under Manage memories and the original conversation, plus any files holding the same detail.
05Can I see what ChatGPT inferred about me?
Partly. There is a memory page where you can view, add, correct, or dismiss what ChatGPT may use to personalize replies. Tech Times reported the new architecture offers only a limited audit trail, so you see the current state but not always when or why it changed.
The takeaway
As of June 2026, ChatGPT builds an inferred, self-revising profile of you by default, and that capability is reaching free accounts. Turning memory off and stopping model training are two different switches in two different menus, and deleting a chat does not delete the memory it produced. Decide deliberately: leave it on for casual use, or turn both off and use temporary chats for anything you would not hand a vendor to keep and learn from.
