In CM systems, controllers can learn how to use different parts of the model network to effectively solve new problems by reducing its search space. Even though the prediction is only for the next time step, the system can learn by memorizing important events that happened a million steps ago.