Product1 publisher3 min readPublished
ChatGPT now runs every answer through a profile it updates on its own
Wired's walkthrough of the memory settings shows ChatGPT deciding for itself what to store about you and carrying it into every reply. Run one prompt on two accounts and you are running two different tests.
The Product Desk · Product desk
What happened
- Wired reports that every ChatGPT answer now arrives with memory context attached, drawn from what the app has stored about the user across past interactions.
- An update in June introduced what OpenAI calls dreaming, which edits memories over time and sifts them automatically, leaving the model to decide what is worth keeping.
- Turning memory off leaves stored memories on file unreferenced, and a full delete leaves past conversations, which get indexed again if memory is switched back on.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint A prompt comparison run on two consumer accounts is not a controlled test, so a team arguing about output quality cannot settle the argument inside the product it is arguing about.
- decision Whoever owns the rollout now has to choose a memory posture per workflow, because the default keeps writing a profile that sits unread by the team.
- exposure A wrong inference about where someone lives or what they do for a living shapes customer-facing drafts until a human opens Personalization and reads the summaries.
- cost Cleanup is per-line handwork, and the person doing it pays twice: re-enabling memory sends ChatGPT back through the old conversations.
Spend a week reading up on how to replace a light switch and ChatGPT starts treating you as a DIY enthusiast, Wired reports, and look up places near somewhere you are visiting and it can decide you live there [2]. To see what it has settled on, you click your profile icon, choose Personalization, scroll to the Memory section, then hit Manage next to Memory summaries [10][11].
The April 2024 version left that judgement with the user. You said "remember this" or the detail was lost in the flow of conversation [3]. Those early memories were fixed, so a party you had mentioned stayed in the future after the date passed [4]. The June update, which OpenAI calls dreaming, made memories editable over time and handed the sorting to the model, so no manual request is needed [5]. Wired summarises OpenAI's three aims for the feature as carrying forward useful context, following preferences and constraints, and staying current over time [7].
A team gets something different out of this than a solo user does. Two colleagues typing the same prompt are typing it against two different stored profiles, and Wired reports that all answers now come with memory context [1]. Neither profile holds still, because ChatGPT refines what it thinks it knows at every interaction and, in Wired's words, "won't always make the right assessments" [6].
Correcting it is handwork. In the summaries you can highlight a line and pick Make a correction or Don't mention this again, or type an instruction such as "forget that I'm a journalist" into the prompt box [11]. A full reset is the three dots, then Delete and turn off memory, then a confirmation [12]. Counting the web app steps, that is four actions to read the profile and three more to clear it, seven in all [15]. The toggle does less than the label suggests: switch memory off and existing memories stay on file, unreferenced [9]. Delete them and your past conversations remain, so turning memory back on indexes those chats all over again [12].
The sorting question for Monday is whether a piece of work has to give the same output to two different people. Benchmarking a prompt does, and so does anything a customer reads over your name. That work does not belong on an account that is quietly building a profile. Choosing dinner, or picking up a project you left half-finished, is the other case, and memory helps there [7].
Wired says nothing in a temporary chat is committed to memory and that the chat is forgotten as soon as it is closed [14]. Whether existing memories still inform the replies inside one, the walkthrough does not say [17]. For evaluation work that difference decides whether a temporary chat gives you a clean baseline or only stops new writes.
What to watch
- An admin control that sets memory off across every seat in an account, instead of one profile at a time.
- Any OpenAI change that shows which stored memory shaped a particular answer.
- Exportable memory summaries, which would let a team compare two accounts' profiles before blaming the prompt.