Memory and new controls for ChatGPT

OpenAI just started rolling out a feature that's been the obvious next step for a while now: ChatGPT can remember things about you across conversations. Mention that you're vegetarian, or that you're building a side project in Rust, and the bot can carry that context into next week's chat without you re-explaining yourself every single time.

The framing from OpenAI is careful, almost cautious. This is a test, not a full launch, and the company is leaning hard on the word "control." You can tell ChatGPT to forget something specific, wipe its memory entirely, or turn the whole feature off if you'd rather it stayed forgetful. That's a deliberate response to the criticism every AI company eventually faces: people want convenience, but they also don't want a chatbot quietly building a dossier on them.

What's notable here isn't the technology itself — persistent memory for language models isn't new, plenty of startups have shipped versions of it. It's that OpenAI is doing it at ChatGPT's scale, with hundreds of millions of users, which means the privacy stakes and the potential for weird, unintended recall are much higher. Remembering a user's coding preferences is harmless. Remembering something they mentioned in a low moment six months ago and bringing it up unprompted is a different kind of problem entirely.

OpenAI says memory will roll out gradually and that user settings will determine how much the model retains. No word yet on how long memories persist by default, or whether enterprise and API customers get a different set of rules than the free consumer product. Those details matter more than the headline feature, honestly, because the difference between "helpful assistant" and "assistant that remembers too much" is entirely in the fine print.