I remember a thread from someone who was trying to keep the assistant in character, and how they got better results using davinci than gpt-3.5-turbo. They later realized that message order apparently mattered. If the personality was always appended in a system message as the very last message, it would mostly adhere to it. Maybe that helps.