Reading time: 11 minutes
ChatGPT memory keeps preferences and details from your chats and uses them to personalize answers, searches and shopping suggestions. As a result, the AI can recommend different brands to each person. This ChatGPT personalization applies only to the account that uses it, and you can turn off ChatGPT memory in the settings.
In short
- ChatGPT memory combines past chats, saved memories, custom instructions and, depending on the plan, connected apps such as Gmail.
- According to OpenAI, memory can change the query sent to search and the products suggested. In other words, ChatGPT personalization also shifts the brands in the answer.
- In an audit by Unusual, changing the profile of the person asking changed up to 75% of mid-market brand picks. For category leaders, however, the change stayed between 20% and 29%.
- To test your brand, it is not enough to turn off ChatGPT memory, because custom instructions still apply. So use an Unpersonalized temporary chat.
What does ChatGPT memory store?
ChatGPT memory stores preferences and details the AI finds useful, not every sentence of your chats. According to OpenAI, the sources can include past chats, saved memories, custom instructions, files in your Library and connected apps such as Gmail. This varies by plan and region.
Also, ChatGPT does not check these sources for every question. It only does so when personal context is likely to improve the answer. In that case, Sources may appear below the answer and show the memory or chat it used.
Much of this memory is created without a request. In a study led by the Max Planck Institute for Software Systems, researchers analyzed 2,050 saved memories from 80 users, mostly in the US and Europe. Only 4% came from a user request. For the other 96%, ChatGPT itself decided to save the information.
However, ChatGPT memory serves to personalize your account, not to train the model. Even so, OpenAI may use chats and memories from personal accounts to improve models if Improve the model for everyone is on. The difference between training and context is in our guide on how ChatGPT works. So telling ChatGPT about a launch does not change what your customers see, as our guide on why ChatGPT is outdated explains.
Does ChatGPT remember your conversations?
Yes, when Reference chat history is on. In that case, it uses relevant details from past chats, and what it remembers changes over time. Saved memories, on the other hand, are stored separately. That is why deleting the original chat does not erase the memory that came from it.
How to see what ChatGPT knows about you
Ask ChatGPT what it remembers about you, as OpenAI suggests. Then open Settings, Personalization and Memory to read the memory summary. But the summary is an overview and may leave out details, so the direct question completes the picture.
Does ChatGPT memory change the brands it recommends?
It can. According to OpenAI, the products ChatGPT shows take into account your question and context, such as memory and custom instructions. In addition, memory can feed into the query it sends to web search. So two people with the same question can get different brands.
OpenAI’s example is a search for dog costumes. If the person has said they dislike clowns, ChatGPT may drop clown costumes from the list. By the same logic, we infer that a preference for local brands or a tight budget can also change the next shopping list. In shopping research, ChatGPT also uses past chats and memory to build a personalized buyer’s guide, according to the company.
When ChatGPT searches the web, memory can add a saved detail to the query, such as a city or a food preference. In other words, the list of pages is already different for each person.
ChatGPT personalization also reaches the ad below the answer. According to the release notes, with ads personalization on, OpenAI also uses past chats and memories to pick the ad. It does not share this data with advertisers.
What tests show about profiles and brands
We found no public study that measures the effect of real ChatGPT memory on brands. The closest one simulates memory with a sentence about who is asking. Unusual, a company that sells consulting to help brands get picked by AI agents, ran 8 shopping questions with 10 profiles, in 2,000 runs. The profiles included a solo founder in the US and a small business owner in London.
Changing the profile changed the brand list more than repeating the question with the same profile. For category leaders, only 20% to 29% of the picks changed. For mid-market brands, however, the change reached 75%. Still, the test came out as a preprint, without peer review. It also used questions in English, two models (from OpenAI and Anthropic) with fixed settings, and a single day.
How does location change ChatGPT’s answer?
ChatGPT estimates your region from your IP address, such as country, state or city, and uses it for local results. Device location, which is more precise, is optional and off by default. Memory, on the other hand, can keep a city you mentioned in a chat and carry it into search.
For brands, this matters for local questions. According to OpenAI, a VPN or your network can change the region estimated from your IP. Therefore, someone testing in São Paulo may not see what a customer in Recife sees (both cities in Brazil). The company even suggests adding your city, neighborhood or postal code (CEP, in Brazil) to the question when results show the wrong region.
In the Unusual test, by the way, regional brands only appeared when the profile named the region. What ChatGPT passes on to partner search engines is in our guide on how to show up in ChatGPT.
Does temporary chat turn off ChatGPT memory?
Yes, if you choose Unpersonalized before the first message, the default since August 2026. In this mode, temporary chat does not use memory, custom instructions or plugins. Personalized mode, on the other hand, lets ChatGPT personalization use your existing memories. Neither mode creates new memories or adds the chat to your history.
The choice lasts for the whole chat, because you cannot change it after the start. OpenAI may keep a copy for up to 30 days for safety. And if you save the chat, it becomes a regular chat and can feed ChatGPT personalization in future answers.
However, Unpersonalized mode does not make every answer the same. The temporary chat page does not mention location, which comes from your IP. The model also changes by plan, and OpenAI warns that the same question can produce different answers.
How to turn off ChatGPT memory
To turn off ChatGPT memory, open Settings, Personalization and Memory. Depending on your account, you will find the Memory, Reference saved memories and Reference chat history toggles there. Turning off Reference saved memories also turns off chat history. However, if you turn off ChatGPT memory, your chats are not deleted, and each control has its own effect:
- All memory: turn off Memory. To erase the summary, use Delete and turn off memory in the three-dot menu.
- Chat history: turn off Reference chat history. What came from your chats leaves OpenAI’s systems within 30 days, and saved memories stay.
- One detail: ask ChatGPT to forget it or delete the item in saved memories. Also delete the chat where the information came up.
- Instructions: in Custom Instructions, turn off Enable customization or delete the text, which applies to all chats.
- One-off chat: use an Unpersonalized temporary chat.
If you turn off ChatGPT memory and turn it back on later, it can create new memories from the chats left in your history. OpenAI may also keep records of deleted memories for up to 30 days, for safety and debugging.
How to test your brand without ChatGPT memory getting in the way?
Test your brand outside your own ChatGPT personalization, in an Unpersonalized temporary chat or while logged out. After all, people who work on a brand talk about it with ChatGPT, and memory can keep that context. As a result, your test may show the brand because of your account, not because of what customers see.
Just turning off ChatGPT memory does not solve this, because custom instructions still apply while customization is on. Then follow four steps:
- Fix the region: test from the customer’s city or add the city to the question.
- Repeat each question: the list changes with every run, so count how often the brand appears, not a single screenshot.
- Create customer profiles: repeat the main questions with a context sentence, such as “I prefer Brazilian brands.”
- Compare the profiles: see which segments include the brand and which ones drop it.
The profile sentence does not copy anyone’s real memory, but it shows which segments move the list. The authors of the Unusual audit draw the same lesson for anyone who measures:
“Any measurement of AI brand perception must condition on the buyer persona supplying the query.”
Will Jack et al., Unusual (2026)
For the full protocol, with the number of runs and margin of error, see our guide on how to measure brand mentions and citations. The quick test is in our guide on how to check your AI visibility.
What does ChatGPT memory change for brands?
It changes the target. With ChatGPT personalization, each account may see its own list, so a brand has no single position. Also, a brand cannot get into anyone’s memory on its own. But it can match the context that memory brings to the question, such as city, budget and preferences.
- Write down your attributes: cities and neighborhoods served, price range, national origin and restrictions, such as gluten-free. In our view, when memory feeds into the query, a page with this data has a better chance of matching the search.
- Create pages by segment: “best X for small businesses” or “for clinics in Curitiba.” In the Unusual audit, more specific profiles concentrated the picks into smaller lists.
- Take care of product data and reviews: when shopping, ChatGPT matches the person’s context with product metadata and third-party content, such as reviews.
- Measure by segment: use your presence without ChatGPT personalization as the baseline, and track questions with profiles separately.
What we found when analyzing 9 Portuguese pages about ChatGPT memory
We analyzed the 9 Brazilian Portuguese pages that showed up most in 8 searches on the topic. One of them was “o que o chatgpt sabe sobre mim” (what does ChatGPT know about me). Five are from 2024 or 2025. Only 3 warn that deleting a chat does not erase the saved memory, and only 2 separate saved memories from chat history.
None says that memory can change web search or explains location by IP. Only 1 gives a shopping example: a travel plugin that would suggest your favorite airline, in 2024. Of the 6 that mention temporary chat, none warns that it can use memory in Personalized mode. Finally, only 1 links to OpenAI, and it is the 2024 announcement.
Frequently asked questions
It can, but Plus and Pro plans have automatic management to prevent that. According to OpenAI, ChatGPT keeps the most relevant saved memories at the top and moves the others to the background. With improved memory, these plans also got twice the capacity.
Yes. On the free plan, ChatGPT memory uses saved memories and references recent chats, according to OpenAI. Library files and connected Gmail, however, only count as sources on Plus and Pro, where the feature is available. Details vary by plan and region.
Based on what OpenAI documents, only if the purchase went through the person’s ChatGPT account, such as a chat about the order or a saved preference. After all, a purchase on your site alone is not among the memory sources OpenAI lists. On Plus and Pro, however, a connected Gmail account can also count. What your store controls is in our guide on GEO for e-commerce.
Want to know how your brand shows up in AI answers today?
NotHuman is opening spots for agencies and brands to track how they appear in AI answers. Join the waitlist and we’ll let you know when a spot opens up for you.
Main sources
- OpenAI, Memory in ChatGPT (2026).
- OpenAI, Temporary chat in ChatGPT (2026).
- OpenAI, Shopping with ChatGPT Search (2026).
- OpenAI, Searching the web with ChatGPT (2026).
- OpenAI, ChatGPT Custom Instructions (2026).
- OpenAI, ChatGPT release notes (2026).
- Unusual, audit of how the asker’s profile affects brands recommended by AI (2026).
- Max Planck Institute for Software Systems, study of 2,050 saved memories from ChatGPT users (2026).
