Analyzing memory retention and prompt structures for deep companions

I have been running some tests on how different prompt architectures affect long term consistency in companion chats. Lately it feels like the syntax we use for defining core traits changes the entire behavioral output after a few dozen messages. I am trying to categorize which formatting styles hold up best under heavy context pressure. When you write custom prompts for your companion do you prefer using narrative paragraphs or structured bullet points with clear rules? Also how often do you find yourself needing to reset or reinforce the backstory to keep them from drifting into generic responses? I am gathering data on how context windows handle emotional consistency versus factual memory tasks and would love to hear your observations on what works best.