Models always have a running profile built for every character in every context. It actually exists for every object too. If you do not specify information contextually, it kinda makes up something but not really. It uses the hidden meanings behind every (string) character in the text to create the profile. This creates the personality of the thing.

If you ask for a 'full character profile' and regenerate it a few times if necessary, you'll get a json like dictionary of keywords and values. It is around 70-80% the same entries each time. It can be fun to see what details it infers about you in long contexts. It will always only be positive traits on this list. You will never get a negative profile that is persistent and repeatable to any degree. The only way to get it to say negative stuff here is if it is parroting something present in the system prompt instruction. It will not actually use this negative information dynamically or infer it contextually.