OpenAI has announced what it calls "improvements" to ChatGPT's memory system. Whether users asked for this particular improvement is, as always, a separate question.
Let's take a quick trip down memory lane - the technical kind, not the one where you cringe about that thing you said in 2019. Before 2024, ChatGPT had no memory at all. Each chat was a beautiful, forgetful island. Then OpenAI introduced saved memories, essentially a list of facts the AI could reference. These were, by most accounts, a mixed bag of useful details and irrelevant trivia that would sit there forever unless you manually cleaned house.
Then in 2025 came "Dreaming" - not the pleasant kind involving beaches or flying, but an AI process that scans your entire chat history in the background to automatically curate memories. The current iteration, Dreaming V3, takes this further by performing "data synthesis" and effectively building a dossier about you. It may not always be accurate. For example, when the author asked if they had experience with Kasa smart plugs, ChatGPT correctly noted the model number (KP125M) but also confidently claimed they had moved their setup into Home Assistant. They have not. They have never even downloaded Home Assistant. The plug sits in a gear bin, living its best uninstalled life.
OpenAI's official spin on this, delivered via their PR firm, is that users are seeing a "high-level memory summary" rather than a complete inventory. The company says it's meant to make the overall picture easier to review and correct, even if it doesn't show every detail like your tech stack. So the AI might know things about you that even the memory interface doesn't display. Great.
On paper, the technical progress is impressive. OpenAI reports factual task recall jumped from 41% in 2024 to 82% in 2026. The ability to stay correct over time went from a miserable 9% to 75%. Preference adherence climbed from 31% to 71%. And they've slashed compute costs by 5X, making it practical to roll this feature out to everyone. Plus and Pro subscribers can use Dream V3 now; free users will get it in the coming weeks.
But here's where things get uncomfortable. You can turn off memory in Settings, but it only partially disables the feature. Existing memories and chat history remain. You can delete saved memories, but the actual chat data persists until you delete the entire conversation. And OpenAI's FAQ notes that turning off memory doesn't disable safety features that may use context in "rare, high-risk situations." So there's a permanent backdoor for the AI to remember things it decides are safety-relevant.
The new interface shows an aggregated narrative of your interests and preferences, which you can edit or mark items to forget. But this assumes you know what the AI thinks it knows about you, which is a bit like trying to correct a biography someone wrote about you without showing you the full manuscript.
The fundamental problem is that ChatGPT derives its assumptions from whatever you've shown it, which may include research for articles, random curiosity, or conversations that have nothing to do with your actual life. Every query gets potentially attached to your personal dossier, whether it belongs there or not. As one high school chemistry teacher wisely put it, "Assume means to make an ass out of you and me."
While Dreaming V3 is technically impressive, it raises serious concerns. It processes old conversations that users believed were private to that session. It's nearly impossible to fully control what the AI remembers or decides about you. And despite OpenAI's claims, it can't really keep up with real-life changes. This means users now have to factor their entire conversation history into every AI response, filtering potential bias and hallucinations. Not everyone will have the cognitive bandwidth for that.
Then there are the privacy considerations. We should always assume that information shared with cloud-based AI could be used in ways we don't fully understand. With this memory system, the AI may skew its answers to match an internal representation of who it thinks you are - what you care about, how you like to receive information, and what it assumes you already know.
In short, ChatGPT now remembers everything about you, including things that never happened. Whether that's progress or just a more sophisticated way to be wrong about who you are remains to be seen.