I'm very interested in this too. I feel like when my Claude session compacts I immediately have to re-alert it to critical aspects of the task, but it should be possible even to have a secondary low-skill agent crawl over the whole context window semi-continuously and nominate pieces for removal or summarization. Or at a higher level, have something able to decompose, like hey I realised we're actually working on three distinct aspects of this problem, let's split the context into those three and I'll just retain a high level summary of what's in the other ones, enough to know if I should pass off control to one of the others for a particular ask.
I feel like this is kind of subagents, but it doesn't have the clear task -> work -> return result -> die model, rather it's more closely aligned to how I myself multitask, maintaining several long-lived threads of execution, and having to intentionally "forget" about the other ones when it's time to lock in and do real work on the most important one.
yeah ive read so many takes on how people survive compaction amnesia ive wired up my own "protocols" that are okay, have used things like beads, am okay burning a buttload of context right at the beginning of a post-compaction just to get up to speed and oriented.. but i still end up with a decent amount of sprawl and lately the writing style of claude has just gotten so dense and nonsensical that my eyes are starting to glaze over.
this seems to be the fate with all the big grand 'god' projects people are doing with AI. super epic omega knowledge bases, super epic omega personal platforms, etc. I don't think enough people admit that whatever they're working on has evolved into something they don't fully understand.
I have run in to the same thing. Most recently I told my pi-agent to fork a git repo since our work had morphed into two different tasks that shouldn't be packed into the same context and workflow. Git fork is what I will use as a tool for this division in the future.