Analyzing memory retention limits and context windows in long form companion chats

Hey everyone. I am currently running some tests on how different platforms handle long term memory during extended roleplay sessions. Specifically I am looking at when the model starts to forget earlier setup details and how it handles contradiction handling when the user updates the backstory. If you have done any deep testing on context degradation or noticed specific patterns in how your companion recalls old events please share your observations. Do you find that periodic summaries help maintain consistency or does it just confuse the local cache further? I am trying to map out a reliable way to keep multi week narratives stable without constantly reminding the system who it is supposed to be. Any technical insights or systematic approaches you use would be greatly appreciated for my notes.