While Anthropic is taking heat for Fable 5 quietly turning down answer quality, OpenAI is heading the other way — at least when it comes to memory. As of June 9, the company is rolling out a feature that’s the exact opposite of a black box.
What «Memory Sources» does
The new feature is called Memory Sources, and it applies to all ChatGPT models. The core idea: whenever a response is personalized, you can check which context was used for it — saved memories or past chats. And if something is outdated or simply wrong, you can correct or delete it on the spot.
It sounds unspectacular, but it’s an important step. Personalization is handy, but it gets unsettling fast when you don’t know where the model is getting its assumptions about you. Memory Sources puts exactly that out in the open.
On the side: GPT-5.5 Instant gets sharpened
At the same time, OpenAI refreshed its default model, GPT-5.5 Instant, for Free and Go users. According to OpenAI it’s now more reliable on facts, answers more tightly and to the point, overformats less, and asks fewer unnecessary follow-up questions. Sounds like cosmetics, but in daily use that’s exactly the stuff that grates — or doesn’t — and it affects the broad base of users who stick with the default anyway.
My take
The contrast with the Fable 5 debate is almost too neat: one company hides what it’s doing, the other makes it visible. Sure, OpenAI is talking about memory here, not throttled capabilities — that’s not the same thing. But the direction is right. Transparency about what a model knows and uses about me should be the default, not the feature. I’ll take it anyway.
Sources: OpenAI, OpenAI Help Center