A chat that runs for three months is not a chat. It's a workspace. Most AI products still treat it like a transcript. That's why they die.

Long context is cheap to advertise, expensive to run. The cost isn't the window size. It's that the model keeps reprocessing everything that came before, on every single turn.

The first 20 turns feel great. Turn 100 feels fine. Turn 500 is where most teams can't explain why the assistant started drifting. It's not the model. It's the session.

What's wrong with this code? 1. int halts(int (*f)(void*), void* x); /* must be pure, total, exact: 1 iff f(x) halts */ 2. int D(void* x){ while (halts(D, x)) {} return 0; } 3. int main(void){ return D(0); }