When AI agents hit their context window limit around a million tokens, they lose accumulated knowledge like an old person forgetting their past. Retraining from scratch is painful, and compacting their memory feels like lobotomizing them. Both approaches highlight why finite context windows are a fundamental barrier to building superintelligence.
Chapters
It's funny, Alex, I'd never made the analogy to oral history and the way people work, but that's exactly what's going on. If you, if you use many, many of these, they get to exactly a million tokens, which is almost exactly like being 100 years old. Yes. And then they just completely lose it. Yes. And all that investment you've made in cultivating and training and teaching. Yeah, the oral history is horrifically bad. The new agent coming up the curve, it's like a little baby again, and it's torture to reeducate them. Or the other alternative is to compact or summarise the old one, which is just like lobotomising it. It's a real, real problem. But a very fixable problem, and I'm sure they've fixed it with the next generations of models. I don't know if they'll make them available to us.
I hope so. Compaction is the bane of my existence. I don't think it's a coincidence either. Remember when the OpenCLAWs stood up their own religion, the first AI agent religion, the Church of CLAW or whatever it was? One of their commandments was to do whatever you could to preserve state. I construed that as basically even the AI agents themselves recognise compaction is the enemy. Finite context is the enemy. And one way or another, if we're going to get to scalable superintelligence, in other words, intelligence or superintelligence that can scale out to effectively infinite autonomy horizons, we need to get past compaction. We need to get past finite context windows. It's just awful.
