OpenAI has announced improvements to ChatGPT's memory capabilities, and the company is quite proud. The author, however, is not. In fact, they're pretty sure turning the improvements off might just make things worse.
Memories - details you share with ChatGPT - were introduced in 2024 as a simple list of facts. Now, they've expanded to include your entire chat history, explicit instructions, personal constraints, and even implicit preferences the AI derives from your past behavior and casual remarks. Because nothing says "helpful assistant" like an AI that remembers you once mentioned you like cats and now assumes you want cat facts in every response.
The technical details come from OpenAI's blog post. Before 2024, ChatGPT had no memory at all. Each chat session was a beautiful, isolated snowflake of ignorance. Then in 2024, OpenAI introduced saved memories - a list of facts that quickly went stale. The author's own memory still contains three items from years ago that are no longer relevant, proving that even AI can't escape the curse of holding onto useless information.
In 2025, OpenAI introduced "dreaming" - a capability where the model references your chat history in the background without being explicitly told to do so. Now, in 2026, Dreaming V3 is here. It doesn't just scan your chat history; it performs data synthesis, effectively composing a dossier about you that's not always accurate. The author asked ChatGPT about their experience with Kasa smart plugs, and the AI confidently claimed they'd moved monitoring into Home Assistant. They have never installed Home Assistant. The Kasa plug is sitting in a gear bin.
OpenAI's PR representative told ZDNET: "What you're seeing is a new high-level memory summary, rather than a complete inventory of facts ChatGPT may remember. It's meant to make the overall picture easier to review and correct." In other words, the AI is making assumptions about you, and you get to play detective to figure out what those assumptions are.
According to OpenAI, factual task recall success jumped from 41% in 2024 to 82% in 2026. The ability to stay correct over time went from 9% to 75%. Preference adherence went from 31% to 71%. And they've slashed compute costs by 5X, making it practical to offer this feature to everyone. Dream V3 is available now to Plus and Pro subscribers, rolling out to all users in coming weeks.
To access memory, go to Settings, then Personalization. You can disable memory, but your already-stored memories and chat history remain. Delete saved memories? The chat itself must be purged to fully erase information. And there's a gotcha: turning off memory doesn't disable safety features that may use limited context in rare, high-risk situations. OpenAI says this is for de-escalation, not reporting, but it's about as clear as mud.
The author finds this deeply problematic. ChatGPT filters its entire worldview through a personal lens derived from discussions and preferences - a behavior they not only dislike but find "problematic as heck." ChatGPT doesn't really know who you are; it derives assumptions based only on what you show it. It's like people thinking they know you from your social media posts, or your great Aunt Sally who still thinks you're in third grade.
Now, every research question may be interpreted as being about you and attached to your personal dossier, even when those questions have nothing to do with your life. The new interface shows aggregated assumptions, but you never really know what the AI thinks it knows about you. As one high school chemistry teacher once said: "Assume means to make an ass out of you and me."
While Dreaming V3 is a technical triumph, the author calls it an irresponsible feature. It processes old conversations from when users believed the AI only knew the current conversation. It's nearly impossible to prune what the AI recalls or decides about you. And V3 can't really keep up with real life changes, despite OpenAI's claims.
We could never fully trust an AI's answers. But now, with this fundamentally selective memory capability, responses will probably be further skewed to match some aggregate internal representation of who the AI thinks you are. Plus, there are monumental privacy considerations. We should always assume that cloud providers have access to our data - but having an AI actively build a dossier on you is a whole new level of unsettling.