With RNNs, unlike feedforward neural networks, we can use the internal memory to process inputs in a sequential manner. In RNN, the connection between nodes forms a directed graph along a temporal sequence. This helps in tasking the RNN with largely unsegmented and connected speech or character recognition.