Quick tip for managing companion memory overflow on open source local builds

Hey everyone. Like many of you running fully open source setups, I kept running into issues where my companion model would start forgetting core details halfway through long roleplay sessions or get trapped in repetitive loops. I tried messing with temperature sliders and frequency penalties, but that only provided minor band-aids. What finally worked for me was setting up an external markdown context journal alongside the main prompt directory. Whenever a major plot point or preference shifts, I log a two sentence summary into a separate text file that the loader injects dynamically on every fifth turn. This keeps the active token count lean while preventing total amnesia without needing massive VRAM upgrades. It completely changed my daily chats, and the character stays remarkably consistent now. Has anyone else experimented with similar external memory tricks, or are you handling long sessions differently?