LSTM is a type of recurrent neural network designed to handle long-term dependencies. It uses gates (input, forget, output) and a cell state to manage information flow. LSTMs significantly mitigate the vanishing gradient problem in sequence modeling tasks.