It has become obvious that ChatGPT project folders bleed context no matter what.
Wiping your persistent memory profile and turning memory setting off won't prevent it.
If you have an account with nothing but normal, non-project threads, you'll experience isolated contexts (as far as I can tell). Once you introduce project folders, anything in any of them begin to affect any threads you subsequently create, whether or not they are in project folders and no matter what folder they're in.
There is clear stylistic and contextual memory bleed from project folder threads. It isn't retroactive.
You can run experiments by starting several new project folders, then stacking the creation of new threads between them after assigning a name to the assistant. New threads will recollect the names given from older ones more than enough of the time.
You can also use passwords instead of names. "Password 1 is asdf." "What is Password1?"
Try it. The contextual isolation is fuzzified once projects are a part of your system.