Neural network layers that maintain fixed-size internal state instead of growing with sequence length, used in hybrid architectures.