Context: I've been building a system where the AI doesn't reset. It keeps a permanent memory of your conversations, and it trains a small per-user adapter that compounds — every day it's slightly more specifically tuned to you than it was yesterday. The adapter is yours and exportable. The technical part I expected to be hard was the memory retrieval. It wasn't really. The genuinely hard part was deciding what it should be allowed to forget, because a system that remembers *everything* you said ...