LSTM Memory Cell

An LSTM separates a cell state from a hidden state.

The cell state is designed as a more stable memory path. Gates decide how much old cell state to keep, how much new candidate information to write, and how much of the cell state to expose as hidden state.

This structure gives gradients a better route across time than a plain RNN in many settings.

Reading the idea

Do not treat the cell state as perfect memory. It is still learned, limited, and trained from data. But it is a more deliberate memory mechanism than a single plain hidden update.

Exercise

If an LSTM has a cell state and a hidden state, how many state vectors are named here?