In 1997 Sepp Hochreiter and Jürgen Schmidhuber publish "Long Short-Term Memory" in Neural Computation, introducing LSTM: a recurrent neural network with gates controlling information retention across arbitrary-length sequences, solving the vanishing-gradient problem. LSTM's three learnable gates (input, forget, output) let information persist across hundreds or thousands of steps, becoming the state of the art for sequence processing for two decades until the Transformer (2017) displaced it.