The OpenAI API unlocked a whole new layer of building for me

That ‘Model as a Processor, Runtime as Memory’ split is the gold standard. It solves the context drift problem and keeps costs linear instead of exponential.

I’m curious about the ‘write back’ phase. When the system updates the durable state after a step, do you use an LLM to ‘compress/summarize’ the findings into the state, or is it more of a structured data update (like updating a JSON manifest of the project)?

This approach basically makes the context window size irrelevant since you’re managing the ‘RAM’ yourself. Brilliant stuff.