ChatGPT's Memory Upgrade Is Quietly Reshaping Your Answers. Here's How to Take Back Control

OpenAI's ChatGPT no longer waits for you to say "remember this." Since a major update rolled out in June, the chatbot has been building and refining a profile of you in the background, drawing on everything from your dietary preferences to the projects you've been working on for months. The feature, which OpenAI calls "dreaming," represents a fundamental shift in how the company's flagship product interacts with its hundreds of millions of users. And most of them have no idea it's happening.

The original memory feature, launched in April 2024, was a blunt instrument. Users had to explicitly instruct the bot to retain a specific detail, and those memories were rigid. If you mentioned an upcoming party, ChatGPT would assume it was still in the future even after the date had passed. The new system, by contrast, is flexible and automatic. Memories are built up, edited, and synthesized over time, with ChatGPT deciding on its own what is worth remembering.

OpenAI says the three guiding objectives are to carry forward useful context, follow user preferences and constraints, and stay current over time. In principle, that should produce more personalized and helpful responses. In practice, it means the chatbot is constantly making judgment calls about who you are and what you care about, and it does not always get those calls right.

A Shadow Profile You Never Asked For

The scale of the algorithmic self-portrait ChatGPT constructs is difficult to overstate. A study published in the proceedings of the ACM Web Conference 2026 analyzed 2,050 memory entries from 80 real-world ChatGPT users and found that 96 percent of memories were created unilaterally by the system, without any explicit user instruction. Roughly 28 percent of those memories contained personal data as defined by GDPR, while 52 percent included psychological insights about the participants. Most were grounded in the actual conversations, but the sheer volume of inference happening behind the scenes raises questions about agency and consent that OpenAI has not fully answered.

The user experience of discovering what ChatGPT has remembered about you is often jarring. A 2026 study presented at the CHI Conference on Human Factors in Computing Systems interviewed 20 ChatGPT users and found that most experienced "negative expectancy violations" after learning what the bot had stored. Participants described the memory as machine-like in its unforgettability and detail, and many expressed a strong need for greater visibility, transparency, and control.

How to See What ChatGPT Knows

The controls exist, but they are buried. In the web app, click your profile icon in the bottom left corner, choose Personalization, and scroll to the Memory section. On mobile, tap the menu button, then your profile icon, then Memory. A toggle switch lets you disable the feature entirely.

Disabling memory does not delete what has already been stored. The chatbot will stop referencing those memories, but it keeps them on file in case you decide to turn the feature back on. To actually review and edit the profile, select Manage next to Memory summaries and go through the list item by item.

This is the part most users skip. The memory summary page is where ChatGPT's assumptions about you become visible, and where incorrect or outdated inferences can be corrected. If the bot has decided you are a DIY enthusiast because you spent a week researching light switch replacement, that assumption is sitting in the summary, quietly shaping every subsequent answer.

The Privacy Trade-Off Nobody Agreed To

OpenAI's privacy policy states that memory is optional and that users can view, edit, or delete stored information. The company notes that ChatGPT is trained not to remember sensitive data like health details unless explicitly asked. But the ACM study found that psychological insights and personal data were being captured at significant rates regardless, suggesting the safeguards are more aspirational than enforced.

There is also the question of what happens in temporary chats. OpenAI says temporary conversations do not create new memories and are not used for model training, but the company retains copies for up to 30 days for safety purposes. If a temporary chat is later saved, it converts to a normal conversation and becomes subject to memory and training settings.

Competitors have taken a more conservative approach. Anthropic's Claude, for instance, does not store topics related to health, race, ethnicity, religious beliefs, politics, or gender identity by default. Users can turn on sensitive topic memory if they choose, and Claude notifies them each time such a topic is saved. OpenAI has no equivalent granular control, which leaves users navigating a binary choice between full memory and no memory at all.

Why This Matters Beyond Convenience

The practical benefits of memory are real. Carrying context across conversations saves time on complex, long-running projects and eliminates the need to re-explain basic preferences. OpenAI's own framing emphasizes continuity and personalization.

But the risks compound over time. A 2026 analysis by ZDNet found that ChatGPT's memory retains outdated information and uses it in responses long after it has become irrelevant. The feature's searchable list of saved facts also misses implicit context and subtext, leading to interpretations that drift from what the user actually meant. In one documented case, a user asked ChatGPT to "make sure the salt is balanced" in a recipe, and the bot interpreted it as a permanent preference for all future recipes.

For a tool that millions of people now use for everything from work tasks to personal advice, the gap between what ChatGPT thinks it knows and what is actually true is not a minor inconvenience. It is a structural flaw in the way the system builds its understanding of you.

The fix is not to abandon memory altogether. It is to treat the memory summary page as a regular part of your ChatGPT routine, checking it the way you would check any other profile that shapes how a system treats you. The chatbot is building a version of you whether you engage with it or not. The least you can do is make sure it is accurate.