I’ve been using ChatGPT with a set of explicit preferences for how I want replies generated, and I keep running into the same problem: the model remembers those preferences, but does not reliably follow them.
These are not vague style requests. They are clear standing instructions about things like formatting, verification, structure, level of detail, and how corrections should be handled.
What is frustrating is that ChatGPT often knows exactly what the rule is. If I point out a violation, it can repeat the preference back to me correctly. But on a later response, it may violate the same rule again.
So the issue is not simply memory. The issue is that remembered preferences do not seem to function as hard constraints during response generation.
If a user has explicitly said that something should always be done, or should never be done, that instruction should be applied before the response is produced rather than treated as optional context.
I would like to see a way to make selected user preferences genuinely persistent and enforceable at generation time. Otherwise, users end up repeatedly correcting behavior that ChatGPT already knows they do not want.