Inference needs memory: how context is becoming AI infrastructure