Tip: Handling conversation compaction gracefully
When a Cowork session runs long enough, the system will compress (compact) the earlier part of the conversation to free up space. You’ll see a message like “Compressing our conversation so we can keep chatting.” This is normal and keeps things running, but there’s a catch: after compaction, your assistant is working from a summary of the conversation rather than the full thing. It still knows your CLAUDE.md exists, but it hasn’t freshly re-read it — so it might drift from your instructions or miss skill routing.
The fix: Add a section to your CLAUDE.md like this:
After Compaction
If this conversation has been compacted (compressed to free up context), re-read this file and Skills/_skill-index.md before proceeding with any requests.
——
This works because CLAUDE.md is loaded into the system context on every turn. After compaction, the assistant sees this instruction in its context and knows to go re-read your config and skill index from disk before doing anything else. It’s not a guarantee, but it’s the best self-healing mechanism available right now.
I also considered building a “refresh” skill for this, but decided against it — a skill only runs if someone triggers it, which doesn’t help if compaction happens while you’re away. The CLAUDE.md instruction is passive and doesn’t require anyone to remember to do anything.
I finally got keyboard maestro working so I have to set up David’s idea of periodically telling the chatbot to re read Claude.md and the skills list.
Somebody in the discord mentioned sending the re-read Claude.md and skills list every 5-10 prompts. The thing that worries me about that is completely filling my knowledge window with multiple copies of Claude.md and skills, and so less space for other things.
What have you done to,stave off Claude amnesia?