Viral 'empty room' trend reveals what ChatGPT remembers about you

What's going on

A new trend is spreading on X. You upload a photo of an empty gray room to ChatGPT and add the prompt "Fill this room with what you know about me". The chatbot then generates an image of the furnished room, and every object in it stands for something it has remembered about you.

According to the Chinese site SMZDM, a Portuguese-speaking user posted the image around midday on November 7. Within two days, the post had 2.27 million views and more than 4,400 replies, most of them showing people's own rooms. The topic "ChatGPT Designs Rooms from Users' Chat Histories" showed up on X's trending page. BeInCrypto (via Yahoo Tech) reports that the big crypto account @RoundtableSpace (0xMarioNawfal) also boosted the trend by urging its followers to join in. In China, users are trying the same thing with Doubao, ByteDance's chatbot.

What the images show, and what they don't

SMZDM describes a product manager's room. His current product hangs on the wall, a recruiting agent works through a pipeline on the floor, AI newsletters are piled up, three clocks show the time in Beijing, San Francisco and New York, and the product roadmap runs from the desk down to the floor. The funniest detail, he says, was a giant stamp reading "Remove the AI flavor". He's always asking the chatbot to make his texts sound more human, and it remembered. The only thing in the room that doesn't look like work is a flamingo.

SMZDM stresses, though, that the image doesn't prove how well the AI knows you. It's a sample of what it has stored, and part of it is guesswork. In one test, the only details in the prompt were "in Shanghai" and "the cat's name is Niangao". The image showed the Oriental Pearl Tower outside the window and the cat's name on its bowl. The user had never mentioned a landmark; the AI invented it. So the image mostly shows what's easy to draw, not necessarily what matters about you.

How ChatGPT's memory works

According to OpenAI's Help Center, as quoted by SMZDM, ChatGPT can use more sources than many people realize when memory is on: past chats, saved memories, custom instructions, files in "Library" and "Space", and content from connected apps such as Gmail. Exactly what's available depends on your plan and region. BeInCrypto describes two layers: "saved memories" for things like your name, preferences and goals, and, for Plus and Pro subscribers, access to your chat history even when nothing was explicitly saved.

ZDNet puts this in historical context. In 2024, memory was a list of individual facts that quickly went stale. The author gives examples from his own account: a note about two Google Workspace accounts, a variable from a Scratch project with broccoli sprites, and the fact that he had installed a Kasa smart plug. All of it mattered for exactly one chat, years ago. In 2025, OpenAI introduced a feature called "Dreaming": the model combs through your chat history in the background and organizes memories on its own. It's now on version 3. According to ZDNet, Dreaming V3 doesn't just analyze the data but compiles it into a kind of dossier on you.

OpenAI gives these figures (vendor claims):

The compute cost of this constant background analysis has reportedly dropped fivefold. According to ZDNet, Dreaming V3 is rolling out to Plus and Pro users first and is set to reach everyone, including free accounts, in the coming weeks.

Made-up memories

Not everything went smoothly in ZDNet's test. Asked about his experience with Kasa, ChatGPT named a specific smart plug model that wasn't stored anywhere. It also claimed the author had later moved the setup to Home Assistant, a home automation platform. He says he never installed it, and the plug is sitting in a box. OpenAI told ZDNet that what you see is a new, high-level summary of memory, not a complete list. Some details aren't displayed but may still be used in conversations. According to SMZDM, the Help Center says something similar: the visible summary is only an overview and doesn't necessarily include every detail and source.

Two separate switches: training and memory

From SMZDM's perspective, the key takeaway is that there are two different settings:

• Training: On personal accounts (Free, Plus, Pro), the "Improve the model for everyone" option under Data Controls is usually on. That lets OpenAI use your chats and memories to improve its models. If you don't want that, you have to turn it off yourself.
• Memory: This lives somewhere else; according to ZDNet, under Settings → Personalization → Memory. It decides whether ChatGPT still knows who you are in new chats.

So turning off training doesn't turn off memory. And according to ZDNet, even turning off memory only goes so far: existing memories and your chat history stay put; nothing new gets added. You delete saved memories one by one, but information only disappears completely once you also delete the chats it came from. At the time of testing, the web version already had the new interface; the macOS app didn't yet.

More tips from the sources:

• You can simply ask ChatGPT: "What do you know about me?"
• For one-off questions, there's Temporary Chat. SMZDM points out that order matters: select "no personalization" first, otherwise even a temporary chat can read your existing memories. If you save a temporary chat later, it becomes a regular chat.
• In Doubao, the settings include a "Memory" section where you can delete entries individually. Under privacy and permissions, the "Help improve the model" option is also on by default.
• If you set this up for older relatives, it's best to adjust these settings for them while you're at it.

Risks of sharing the images

If you post your room, take a close look first. Are there landmarks that give away where you live? Are company names, pet names or people's names visible on screens, book spines or mugs? Can you spot anything about family, health, money or travel plans? SMZDM also advises against using an empty-room photo of unknown origin; take your own or generate one instead.

Crypto outlets such as BeInCrypto and The Currency Analytics have a particular warning for crypto holders. A profile that includes your holdings, exchanges or travel plans helps scammers write convincing phishing messages, fake messages designed to steal your login credentials. The reports point to an incident in early 2026 in which developers found wallet data in shared Claude chats that Google had indexed. Binance founder Changpeng Zhao reportedly said the real AI threat to crypto is phishing and data leaks, not cracking Bitcoin's encryption. According to the reports, Ethereum co-founder Vitalik Buterin has built a "paranoid mode" that masks prompts before they reach a model.

Our take

The trend is harmless fun, but it exposes something many people don't know: ChatGPT is increasingly building a profile of you in the background on its own, and with Dreaming V3, free accounts will soon get this too. If you work with AI every day, that's genuinely useful, because you don't have to explain who you are and what you're working on every time. The problems: the feature is on by default, training and memory are controlled in different places, and by OpenAI's own account the visible summary doesn't show everything. ZDNet's test also shows the dossier can contain false information that shapes later answers. ZDNet itself discloses that its parent company, Ziff Davis, is suing OpenAI for copyright infringement. The Dreaming V3 performance figures come solely from OpenAI. Either way, it's worth reviewing your memories and deleting a few entries. That does you more good than just sharing the picture.