The belief was that keeping a session open protected the work. Close it and you lose the thread. Keep it alive and you preserve context. So sessions ran long.
On July 15th I ran /context mid-session to see what I was actually paying for. Conversation messages were 523k tokens. Tool results added another 185k, about 55k of that from whole-file reads when a section would’ve done. Memory, CLAUDE.md, all the skills combined came to 5.5k tokens, cached. The thing I’d been trying to protect was about 1 percent of what I was paying for.
I thought running wrapup at the end of a topic cleared the accumulation. It doesn’t. Wrapup saves to memory. Only opening a new session fully resets the context window. So every time I ran wrapup and kept going, I was adding to a session that didn’t need more.
And closing the session wasn’t losing anything. Transcripts persist on disk. The important durables land in memory. When I open a fresh session, catchup rehydrates all of it.
There was also a story I’d told myself about collisions. I’d blamed past context problems on having multiple sessions open at once, so I kept one long session running all day instead. But those collisions were parallel sessions hitting the same surface simultaneously. Serial sessions, where you close one before opening the next, have no collision risk. I was avoiding the wrong thing.
The rhythm turned out to be simpler. Catchup to orient. One topic. Wrapup when the topic turns. Close. Fresh session. Clean window, everything already behind it.
I also asked my AI to say “good wrap point here” when the thread shifts. Same reflex as the board intake detector. I shouldn’t be the one noticing when to close.
Past me should’ve run /context in week one.