Ah, that makes sense. So, we consider two hidden layers more as "memory" or "buffers", and actually the rule is implemented in just one layer, at least for a single token.