I'm excited about ChatGPT's memory upgrade - but I'm quickly seeing a downside

David Gewirtz/ZDNET

Follow ZDNET: Add america arsenic a preferred source on Google.

ZDNET's cardinal takeaways

• ChatGPT's representation present builds a illustration from past chats.
• Old oregon irrelevant details tin distort aboriginal AI answers.
• Turning representation disconnected whitethorn not afloat erase what ChatGPT knows.

According to a blog station released recently, OpenAI seems rather arrogant of the "improvements" it's made to managing idiosyncratic memories successful the chatbot. I'm not definite I similar them. In fact, I cognize I don't. I'm conscionable not definite whether turning the improvements disconnected volition marque things amended oregon worse.

Also: ChatGPT vs. Gemini's AI representation generation: A azygous punctual tweak makes a large difference

Memories, for the intent of this discussion, are details you stock with ChatGPT. Introduced successful 2024, memories were fundamentally a database of facts the AI could reference. Today, they person expanded considerably to see your full chat history, explicit instructions, idiosyncratic constraints, and adjacent implicit preferences the AI derives from past behaviour and casual remarks.

(Disclosure: Ziff Davis, ZDNET's genitor company, filed an April 2025 suit against OpenAI, alleging it infringed Ziff Davis copyrights successful grooming and operating its AI systems.)

This nonfiction has 3 sections. First, I'll spell implicit the method accusation from OpenAI's recent blog station astir ChatGPT's improved memory capabilities. Second, I'll amusement you the interface elements you tin usage to tweak ChatGPT's memories. Finally, I'll wrapper up with wherefore this diagnostic freaks maine retired and wherefore it mightiness interest you, too.

Let's excavation in.

Dream a small imagination of maine

Before 2024, ChatGPT didn't person memories of immoderate kind. Each chat league lived connected its own, and immoderate you told it was unavailable to immoderate different session.

Also: Compare caller models with our AI Model Release Tracker

Then, successful 2024, OpenAI introduced memories. In the 2024 context, memories were fundamentally a database of facts. ChatGPT inactive has this feature, and overmuch of it is perfectly useless. Here are 3 items inactive successful my ChatGPT memory:

• Has 2 Google Workspace accounts, 1 for A.com (used for message and 57 TB of backups) and different for B.com. User wants to determination A.com email accounts and routing to the B.com Workspace but clasp the A.com Workspace for storage, perchance nether a antithetic domain.
• Has a planetary database called 'veg_positions' successful their Scratch task to way sprite positions, particularly to forestall overlaps among aggregate broccoli clones.
• Has installed the Kasa astute plug.

These items are inactive successful my ChatGPT memory, adjacent though they were lone applicable for the 1 chat league they were discussed in, and adjacent though those sessions were a fewer years back.

This searchable database of saved facts misses implicit discourse and subtext. As you tin see, saved memories went stale fast. Unless curated by the idiosyncratic by hand, determination was nary mechanics to cognize erstwhile the accusation became nary longer useful.

Also: Anthropic launches Opus 4.8, with honesty arsenic its slayer feature

Then, successful 2025, OpenAI began enactment connected what it calls dreaming.

For humans, dreaming helps america process our emotions, consolidate and encode memories, simulate imaginable threats, and reenforce societal and affectional intelligence. A Scientific American article connected the subject down dreaming says, "Dreams assistance modulate postulation connected that fragile span which connects our experiences with our emotions and memories."

My dreams usually impact villagers with pitchforks and torches chasing me, oregon being trapped successful caves with flaming skulls and the occasional IBM mainframe. 

In immoderate case, successful 2025, OpenAI's dreaming capableness meant the exemplary could notation your chat past successful the background, without being explicitly told to look thing up oregon remember. It started to curate memories automatically. ChatGPT's representation operation consisted of saved memories (the 2024 mechanism) positive Dreaming V0.

Also: The 5 myths of the agentic coding apocalypse

Now, successful 2026, ChatGPT's Dreaming diagnostic is astatine V3. Saved memories inactive exist, but they're either replaced oregon augmented with dreaming-based memories.

For example, I conscionable asked ChatGPT, "Do I person immoderate acquisition with Kasa?" It replied, "Yes. You've utilized oregon discussed Kasa KP125M Matter astute plugs, specifically for astute location vigor monitoring." As you saw above, that exemplary fig is not successful the saved memory.

But it's besides wholly wrong. The 2nd portion of the reply to my "experience with Kasa" question was this: "You aboriginal moved that monitoring setup into Home Assistant, truthful your Kasa acquisition appears to beryllium tied to power/energy tracking, not conscionable basal on/off plug control."

Yeah, not truthful much. I person ne'er installed Home Assistant. The Kasa plug is presently sitting successful my cogwheel bin on with a clump of different hardware.

When I reached retired to OpenAI via their PR steadfast astir my concerns, a institution typical told ZDNET, "What you're seeing is simply a caller high-level representation summary, alternatively than a implicit inventory of facts ChatGPT whitethorn remember. It's meant to marque the wide representation easier to reappraisal and correct, but it won't needfully amusement each detail, specified arsenic your tech stack, adjacent erstwhile that discourse tin inactive beryllium utilized successful a applicable conversation."

Today's Dreaming V3 doesn't conscionable scan your chat past successful the background. It performs information synthesis, efficaciously composing a dossier astir you (that's not ever accurate). According to OpenAI, V3 is susceptible of carrying guardant analyzable discourse accusation and tracking dense, multi-session, multilayered long-running projects.

According to the company, factual task callback occurrence jumped from 41% successful 2024 to 82% successful 2026. The quality to enactment close implicit clip went from a 9% taskrate successful 2024 to 75% successful 2026. Preference adherence went from 31% successful 2024 to 71% successful 2026.

Also: Your Claude agents tin 'dream' present - however Anthropic's caller diagnostic works

How does each of this autonomous inheritance processing scale? That's a large portion of the breakthrough that makes Dream V3 possible. OpenAI has slashed the compute outgo by 5X for this level of ongoing analysis. The nonstop effect of this ratio makes it applicable for OpenAI to connection the diagnostic for wide access. The AI is perpetually revising its interior intelligence authorities successful the background, looking astatine timestamps attached to informational nuggets, and efficaciously experiencing passages of clip alongside you (in theory).

The Dream V3 diagnostic is disposable present to Plus and Pro tier subscribers. The diagnostic volition beryllium rolling retired to each users (including escaped users) implicit the coming weeks.

A travel down Memory Lane

I recovered the caller representation features successful the browser mentation of ChatGPT. At the clip of this writing, the MacOS app inactive has the aged representation interface.

To entree memory, spell to Settings, past Personalization. Scroll down to the Memory section.

Screenshot by David Gewirtz/ZDNET

You tin disable the representation capableness (shown astatine 1), but lone partially. If you crook disconnected Enable memory, your already-stored memories volition not beryllium removed, nor volition immoderate of your chat history. ChatGPT won't bash dream-based representation consolidation, but the information volition beryllium there.

Also: Anthropic's Mythos is evolving faster than expected, reports AI information agency

If you privation to delete your memories, spell to the saved memories interface (at 2) and delete saved memories. Even then, each your memories won't beryllium deleted. You'll person to delete the existent chat itself to purge that accusation from the AI's consciousness.

There's besides a weird gotcha. In ChatGPT's memory FAQ, OpenAI says this:

Turning disconnected Memory/Personalization does not disable information features that whitethorn usage limited, safety-relevant discourse successful rare, high-risk situations to assistance ChatGPT respond much safely.

You tin read OpenAI's statement of its information protocols. However, the basal thought is that if a speech implies a information issue, ChatGPT won't delete that accusation and volition usage it to assistance the idiosyncratic de-escalate a situation. It doesn't look similar the AI reports that accusation to anyone, but it's determination (and somewhat unclear).

Finally, you tin deed Manage (at 3) to marque changes to the consolidated illustration ChatGPT has developed astir you.

Also: I tried Google Drive's caller AI cleanup instrumentality to hole 14 years of retention clutter

Let's pat into saved memories archetypal (at 2). Here's an excerpt of mine. As you tin see, they're highly circumstantial (and rather irrelevant oregon wrong). For example, I'm not moving Mint Linux on my Mac Studio, but I was exploring the thought astir a twelvemonth ago.

Screenshot by David Gewirtz/ZDNET

The caller diagnostic is nether the Manage button. Because I don't privation to stock ChatGPT's mentation of my property and preferences, I'm utilizing a screenshot OpenAI provided successful its blog.

Image: OpenAI

As you tin see, it's a communicative astir interests and preferences. You tin prime aspects and people "Don't notation this again" to efficaciously hide that item. You tin besides adhd comments to personalize however ChatGPT groks you.

Thanks for the representation

There are scenarios wherever coagulated AI representation capableness is mission-critical. I've constructed a precise precise acceptable of representation instructions that govern however Claude Code manages my Apple coding projects. I person different representation infrastructure moving connected my OpenClaw location server that allows it to clasp discourse and instructions betwixt tasks.

Also: I utilized Claude Code to vibe codification a Mac app successful 8 hours

But I'm not astatine each comfy with this caller representation mechanics successful ChatGPT. I showed supra however the diagnostic retains wildly out-of-date accusation and past uses this accusation successful its responses. Even Dreaming V3, which supposedly changes with you implicit time, tin beryllium wrong. I showed you the illustration wherever it claims I'm moving Home Assistant, and I've ne'er adjacent downloaded it.

It filters the full worldview done a idiosyncratic lens that's someway derived from discussions and preferences. This is simply a behaviour I not lone dislike but besides find problematic arsenic heck.

ChatGPT doesn't truly cognize who I americium due to the fact that it derives its assumptions based lone connected what I amusement it.

It's similar radical reasoning they cognize maine solely from my societal media posts, oregon my large Aunt Sally, who inactive thinks of maine arsenic a 3rd grader adjacent though I'm a palmy nonrecreational with an precocious degree, oregon an aged buddy who inactive thinks I devour lone junk nutrient and americium incapable of appreciating bully things due to the fact that I erstwhile went done a low-budget convenience signifier backmost successful the 1990s.

Also: This exec offers 4 ways to beryllium a palmy innovator successful the property of agentic AI

ChatGPT present besides tries to respond based connected aged chat conversations, but not each of my conversations are astir me. I'm often researching for articles and projects. These days, each probe question whitethorn beryllium interpreted arsenic astir maine and attached to my idiosyncratic dossier, erstwhile those questions mightiness person thing to bash whatsoever with my life, different than a momentary spot of curiosity.

And portion the caller interface shows its aggregated assumptions astir who I am, I ne'er truly cognize what it thinks it knows astir me, and truthful what assumptions it makes. As 1 of my precocious schoolhouse chemistry teachers erstwhile said, "Assume means to marque an ass retired of you and me."

While Dreaming V3 is intelligibly a method triumph successful breadth, efficiency, and scalability, I'm going to spell truthful acold arsenic to accidental it's an irresponsible feature. First, it processes utilizing aged conversations, galore of which occurred erstwhile the prevailing idiosyncratic position was that the AI lone knew of the existent conversation. Second, it's astir intolerable to prune what the AI recalls oregon decides astir you. And third, V3 can't truly support up with your existent beingness oregon changes, contempt OpenAI's claims that it can.

As overmuch arsenic AI is expected to trim cognitive load connected humans, it really increases our reasoning load erstwhile filtering retired AI bias and hallucinations from responses. Now, we request to origin our full past of conversations with the AI against each answer. But not each users volition person the cognitive subject to analyse and verify the veracity and completeness of each AI response.

Will AI conveniently permission retired accusation due to the fact that its warped presumption of who we are suggests we aren't funny successful an country oregon approach? Will it modify its presumption due to the fact that it thinks we lone privation to person accusation successful a definite format?

We could ne'er truly spot an AI's answers. We knew that going in, backmost astatine the opening of 2023. But now, with this representation capability, with this fundamentally selective representation capability, we tin expect answers volition astir apt beryllium further skewed to lucifer immoderate aggregate interior practice of who it thinks we are, what it thinks we attraction about, and however it thinks we privation to beryllium communicated with.

Also: I had ChatGPT physique maine a escaped PDF exertion due to the fact that I didn't spot it to alteration my files

Then, of course, determination are the monumental privateness considerations. We should ever presume that unreality providers (Google, I'm looking astatine you) store each small item of accusation astir america for aboriginal use. But this level of profiling seems adjacent much invasive and troubling.

Misty water-colored memories

Look astatine this bizarre conversation. I asked, "What is the azygous astir important happening I attraction about?" My answer, unquestionably, is my woman and my small dog. More than thing other successful the world, I emotion those two.

But the AI says, "Preserving your quality to support doing meaningful, autarkic work." Where the hellhole did it get that answer?

Screenshot by David Gewirtz/ZDNET

Yes, I regularly interact with ChatGPT astir the wellness of antithetic foods and aspects of exercise, due to the fact that I attraction astir my health. I besides speech to it astir sustainability questions successful airy of things similar Google's usage of AI answers successful spot of presenting hunt results.

But the astir important thing? Far from it. The AI says it came up with that appraisal due to the fact that of the pursuing conversational factors:

• Web publishing resilience successful the look of AI
• Owned assemblage and newsletter growth
• Effectiveness arsenic a exertion journalist
• Technical independence
• Remote-work viability
• ZATZ Labs and WordPress projects
• YouTube and shaper work
• Home lab, automation, backups, cameras, lighting, and workflow reliability

Note that the AI didn't disclaim that its appraisal was constricted to these conversational factors. It has nary knowledge, for example, that I've been moving connected Mac, iPhone, iPad, and Apple Watch programming projects successful Claude Code. But the specified information that it near Denise and Pixel retired of my "most important" database tells maine everything I request to cognize astir the prime of ChatGPT's assumptions.

They are skewed, inaccurate, retired of date, often highly irrelevant, generalized from precise circumstantial impermanent contexts, and perchance unsafe if utilized to bounds informational responses.

And that's wherefore this caller Dreaming V3 diagnostic freaks maine out. It should rise the hairs connected the backs of your necks, too.

Do you similar an AI adjunct that remembers your past conversations oregon 1 that treats each chat arsenic a cleanable slate? Let america cognize successful the comments below.

You tin travel my day-to-day task updates connected societal media. Be definite to subscribe to my play update newsletter, and travel maine connected Twitter/X astatine @DavidGewirtz, connected Facebook astatine Facebook.com/DavidGewirtz, connected Instagram astatine Instagram.com/DavidGewirtz, connected Bluesky astatine @DavidGewirtz.com, and connected YouTube astatine YouTube.com/DavidGewirtzTV.