Observe ZDNET: Add us as a preferred source on Google.
ZDNET’s key takeaways
- ChatGPT’s reminiscence now builds a profile from previous chats.
- Outdated or irrelevant particulars can distort future AI solutions.
- Turning reminiscence off could not absolutely erase what ChatGPT is aware of.
In line with a weblog publish launched not too long ago, OpenAI appears fairly pleased with the “enhancements” it is made to managing person reminiscences within the chatbot. I am undecided I like them. Actually, I do know I do not. I am simply undecided whether or not turning the enhancements off will make issues higher or worse.
Additionally: ChatGPT vs. Gemini’s AI image generation: A single prompt tweak makes a big difference
Reminiscences, for the aim of this dialogue, are particulars you share with ChatGPT. Launched in 2024, reminiscences have been principally an inventory of info the AI may reference. Right this moment, they’ve expanded significantly to incorporate your total chat historical past, express directions, private constraints, and even implicit preferences the AI derives from previous habits and informal remarks.
(Disclosure: Ziff Davis, ZDNET’s guardian firm, filed an April 2025 lawsuit towards OpenAI, alleging it infringed Ziff Davis copyrights in coaching and working its AI methods.)
This text has three sections. First, I will go over the technical info from OpenAI’s recent blog post about ChatGPT’s improved memory capabilities. Second, I will present you the interface parts you need to use to tweak ChatGPT’s reminiscences. Lastly, I will wrap up with why this function freaks me out and why it would fear you, too.
Let’s dig in.
Dream a bit dream of me
Earlier than 2024, ChatGPT did not have reminiscences of any sort. Every chat session lived by itself, and no matter you instructed it was unavailable to another session.
Additionally: Compare new models with our AI Model Release Tracker
Then, in 2024, OpenAI launched reminiscences. Within the 2024 context, reminiscences have been principally an inventory of info. ChatGPT nonetheless has this function, and far of it’s completely ineffective. Listed here are three objects nonetheless in my ChatGPT reminiscence:
- Has two Google Workspace accounts, one for A.com (used for mail and 57 TB of backups) and one other for B.com. Person needs to maneuver A.com electronic mail accounts and routing to the B.com Workspace however retain the A.com Workspace for storage, probably beneath a unique area.
- Has a world checklist known as ‘veg_positions’ of their Scratch venture to trace sprite positions, particularly to forestall overlaps amongst a number of broccoli clones.
- Has put in the Kasa good plug.
This stuff are nonetheless in my ChatGPT reminiscence, although they have been solely related for the one chat session they have been mentioned in, and although these periods have been a number of years again.
This searchable checklist of saved info misses implicit context and subtext. As you’ll be able to see, saved reminiscences went stale quick. Except curated by the person by hand, there was no mechanism to know when the knowledge grew to become now not helpful.
Additionally: Anthropic launches Opus 4.8, with honesty as its killer feature
Then, in 2025, OpenAI started work on what it calls dreaming.
For people, dreaming helps us course of our feelings, consolidate and encode reminiscences, simulate potential threats, and reinforce social and emotional intelligence. A Scientific American article on the science behind dreaming says, “Goals assist regulate site visitors on that fragile bridge which connects our experiences with our feelings and reminiscences.”
My goals often contain villagers with pitchforks and torches chasing me, or being trapped in caves with flaming skulls and the occasional IBM mainframe.
In any case, in 2025, OpenAI’s dreaming functionality meant the mannequin may reference your chat historical past within the background, with out being explicitly instructed to look one thing up or bear in mind. It began to curate reminiscences mechanically. ChatGPT’s reminiscence construction consisted of saved reminiscences (the 2024 mechanism) plus Dreaming V0.
Additionally: The 5 myths of the agentic coding apocalypse
Now, in 2026, ChatGPT’s Dreaming function is at V3. Saved reminiscences nonetheless exist, however they’re both changed or augmented with dreaming-based reminiscences.
For instance, I simply requested ChatGPT, “Do I’ve any expertise with Kasa?” It replied, “Sure. You’ve got used or mentioned Kasa KP125M Matter good plugs, particularly for good dwelling power monitoring.” As you noticed above, that mannequin quantity just isn’t within the saved reminiscence.
However it’s additionally fully mistaken. The second a part of the reply to my “expertise with Kasa” query was this: “You later moved that monitoring setup into House Assistant, so your Kasa expertise seems to be tied to energy/power monitoring, not simply primary on/off plug management.”
Yeah, not a lot. I’ve by no means put in House Assistant. The Kasa plug is presently sitting in my gear bin together with a bunch of different {hardware}.
After I reached out to OpenAI through their PR agency about my considerations, an organization consultant instructed ZDNET, “What you are seeing is a brand new high-level reminiscence abstract, fairly than a whole stock of info ChatGPT could bear in mind. It is meant to make the general image simpler to overview and proper, however it will not essentially present each element, comparable to your tech stack, even when that context can nonetheless be utilized in a related dialog.”
Right this moment’s Dreaming V3 would not simply scan your chat historical past within the background. It performs knowledge synthesis, successfully composing a file about you (that is not all the time correct). In line with OpenAI, V3 is able to carrying ahead advanced context info and monitoring dense, multi-session, multilayered long-running tasks.
In line with the corporate, factual activity recall success jumped from 41% in 2024 to 82% in 2026. The flexibility to remain right over time went from a 9% taskrate in 2024 to 75% in 2026. Choice adherence went from 31% in 2024 to 71% in 2026.
Additionally: Your Claude agents can ‘dream’ now – how Anthropic’s new feature works
How does all of this autonomous background processing scale? That is an enormous a part of the breakthrough that makes Dream V3 potential. OpenAI has slashed the compute value by 5X for this degree of ongoing evaluation. The direct results of this effectivity makes it sensible for OpenAI to supply the function for mass entry. The AI is continually revising its inner psychological state within the background, taking a look at timestamps connected to informational nuggets, and successfully experiencing passages of time alongside you (in idea).
The Dream V3 function is offered now to Plus and Pro tier subscribers. The function shall be rolling out to all customers (together with free customers) over the approaching weeks.
A visit down Reminiscence Lane
I discovered the brand new reminiscence options within the browser model of ChatGPT. On the time of this writing, the MacOS app nonetheless has the outdated reminiscence interface.
To entry reminiscence, go to Settings, then Personalization. Scroll right down to the Reminiscence part.
You’ll be able to disable the reminiscence functionality (proven at 1), however solely partially. When you flip off Allow reminiscence, your already-stored reminiscences is not going to be eliminated, nor will any of your chat historical past. ChatGPT will not do dream-based reminiscence consolidation, however the knowledge shall be there.
Additionally: Anthropic’s Mythos is evolving faster than expected, reports AI safety agency
If you wish to delete your reminiscences, go to the saved reminiscences interface (at 2) and delete saved reminiscences. Even then, all of your reminiscences will not be deleted. You may must delete the precise chat itself to purge that info from the AI’s consciousness.
There’s additionally a bizarre gotcha. In ChatGPT’s memory FAQ, OpenAI says this:
Turning off Reminiscence/Personalization doesn’t disable security options that will use restricted, safety-relevant context in uncommon, high-risk conditions to assist ChatGPT reply extra safely.
You’ll be able to read OpenAI’s description of its safety protocols. Nonetheless, the fundamental concept is that if a dialog implies a security problem, ChatGPT will not delete that info and can use it to assist the person de-escalate a scenario. It would not seem just like the AI studies that info to anybody, however it’s there (and considerably unclear).
Lastly, you’ll be able to hit Handle (at 3) to make modifications to the consolidated profile ChatGPT has developed about you.
Additionally: I tried Google Drive’s new AI cleanup tool to fix 14 years of storage clutter
Let’s faucet into saved reminiscences first (at 2). Here is an excerpt of mine. As you’ll be able to see, they’re extremely particular (and fairly irrelevant or mistaken). For instance, I am not operating Mint Linux on my Mac Studio, however I used to be exploring the concept a couple of 12 months in the past.
The brand new function is beneath the Handle button. As a result of I do not need to share ChatGPT’s interpretation of my persona and preferences, I am utilizing a screenshot OpenAI offered in its weblog.
As you’ll be able to see, it is a narrative about pursuits and preferences. You’ll be able to choose elements and mark “Do not point out this once more” to successfully neglect that merchandise. You may as well add feedback to personalize how ChatGPT groks you.
Thanks for the reminiscence
There are eventualities the place strong AI reminiscence functionality is mission-critical. I’ve constructed a really exact set of reminiscence directions that govern how Claude Code manages my Apple coding tasks. I’ve one other reminiscence infrastructure operating on my OpenClaw dwelling server that enables it to retain context and directions between duties.
Additionally: I used Claude Code to vibe code a Mac app in 8 hours
However I am by no means comfy with this new reminiscence mechanism in ChatGPT. I confirmed above how the function retains wildly out-of-date info after which makes use of this info in its responses. Even Dreaming V3, which supposedly modifications with you over time, will be mistaken. I confirmed you the instance the place it claims I am operating House Assistant, and I’ve by no means even downloaded it.
It filters your entire worldview by way of a private lens that is one way or the other derived from discussions and preferences. It is a habits I not solely dislike but additionally discover problematic as heck.
ChatGPT would not actually know who I’m as a result of it derives its assumptions primarily based solely on what I present it.
It is like folks considering they know me solely from my social media posts, or my nice Aunt Sally, who nonetheless thinks of me as a 3rd grader although I am a profitable skilled with a sophisticated diploma, or an outdated buddy who nonetheless thinks I eat solely junk meals and am incapable of appreciating good issues as a result of I as soon as went by way of a low-budget comfort part again within the Nineties.
Additionally: This exec offers 4 ways to be a successful innovator in the age of agentic AI
ChatGPT now additionally tries to reply primarily based on outdated chat conversations, however not all of my conversations are about me. I am usually researching for articles and tasks. As of late, each analysis query could also be interpreted as about me and connected to my private file, when these questions might need nothing to do by any means with my life, aside from a momentary little bit of curiosity.
And whereas the brand new interface reveals its aggregated assumptions about who I’m, I by no means actually know what it thinks it is aware of about me, and subsequently what assumptions it makes. As one in every of my highschool chemistry academics as soon as stated, “Assume means to make an ass out of you and me.”
Whereas Dreaming V3 is clearly a technical triumph in breadth, effectivity, and scalability, I will go as far as to say it is an irresponsible function. First, it processes utilizing outdated conversations, a lot of which occurred when the prevailing person perspective was that the AI solely knew of the present dialog. Second, it is almost unattainable to prune what the AI recollects or decides about you. And third, V3 cannot actually sustain along with your actual life or modifications, regardless of OpenAI’s claims that it will possibly.
As a lot as AI is meant to cut back cognitive load on people, it truly will increase our considering burden when filtering out AI bias and hallucinations from responses. Now, we have to issue our total historical past of conversations with the AI towards each reply. However not all customers may have the cognitive self-discipline to look at and confirm the veracity and completeness of each AI response.
Will AI conveniently pass over info as a result of its warped view of who we’re suggests we aren’t serious about an space or strategy? Will it modify its presentation as a result of it thinks we solely need to obtain info in a sure format?
We may by no means actually belief an AI’s solutions. We knew that getting in, again originally of 2023. However now, with this reminiscence functionality, with this essentially selective reminiscence functionality, we will count on solutions will most likely be additional skewed to match some combination inner illustration of who it thinks we’re, what it thinks we care about, and the way it thinks we need to be communicated with.
Additionally: I had ChatGPT build me a free PDF editor because I didn’t trust it to change my files
Then, after all, there are the monumental privateness concerns. We must always all the time assume that cloud suppliers (Google, I am taking a look at you) retailer each little element of details about us for later use. However this degree of profiling appears much more invasive and troubling.
Misty water-colored reminiscences
Take a look at this weird dialog. I requested, “What’s the single most necessary factor I care about?” My reply, unquestionably, is my spouse and my little canine. Greater than the rest on the planet, I like these two.
However the AI says, “Preserving your skill to maintain doing significant, unbiased work.” The place the hell did it get that reply?
Sure, I usually work together with ChatGPT concerning the well being of various meals and elements of train, as a result of I care about my well being. I additionally discuss to it about sustainability questions in mild of issues like Google’s use of AI solutions instead of presenting search outcomes.
However crucial factor? Removed from it. The AI says it got here up with that evaluation due to the next conversational elements:
- Net publishing resilience within the face of AI
- Owned viewers and publication development
- Effectiveness as a expertise journalist
- Technical independence
- Distant-work viability
- ZATZ Labs and WordPress tasks
- YouTube and maker work
- House lab, automation, backups, cameras, lighting, and workflow reliability
Word that the AI did not disclaim that its evaluation was restricted to those conversational elements. It has no data, for instance, that I have been engaged on Mac, iPhone, iPad, and Apple Watch programming tasks in Claude Code. However the mere indisputable fact that it left Denise and Pixel out of my “most necessary” checklist tells me every little thing I must know concerning the high quality of ChatGPT’s assumptions.
They’re skewed, inaccurate, outdated, usually extremely irrelevant, generalized from very particular short-term contexts, and probably harmful if used to restrict informational responses.
And that is why this new Dreaming V3 function freaks me out. It ought to increase the hairs on the backs of your necks, too.
Do you like an AI assistant that remembers your previous conversations or one which treats every chat as a clear slate? Tell us within the feedback under.
You’ll be able to comply with my day-to-day venture updates on social media. You’ll want to subscribe to my weekly update newsletter, and comply with me on Twitter/X at @DavidGewirtz, on Fb at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.
