Good to see someone else going after the stale memory problem. When it catches a memory that’s gone out of date, does it just flag it for you to rewrite, or does it try to work out what the memory should have said and propose that? Propose the corrected memory, flag it, and let me accept or reject it.
Does the handoff prompt do anything different from just pasting the last few messages, or is it carrying state the transcript doesn’t have? I ask because that file should hold the things nobody types: which branch is live, what you tried that didn’t work. The assumption you’re currently standing on is usually nowhere in it. If it’s only a summary of the conversation, a fresh account will read it and still repeat the dead end. What goes in yours that a transcript wouldn’t already tell me?
Your routing section works because it puts the decision in the instructions rather than in the agent’s judgement, which is the only place it can be checked. Most CLAUDE.md files I see list what the agent should do and leave the when to the model. Put the subagent, the trigger and the handoff in one place and they can be reviewed. The rest of the file is a style guide that would survive in any agent’s context.
deleted by creator
deleted by creator
Burn 500M tokens and you learn where they actually went. How did the split look between context you were re-sending and context you were re-reading? The one doing the damage is never the one I suspect. The write-up naming the number is half the reason I want to read it. Good luck with the AMA.
Does the saving come from spacing things out or from keeping each session tighter? Either way it’s the kind of thing worth writing up before anyone else tries it.