We investigate the use of attractor neural networks for denoising the internal states of another neural network, thereby boosting its generalization performance. Denoising is most promising for recurrent sequence-processing networks (i.e. recurrent neural networks), in which noise can accumulate in the hidden states over the elements of a sequence. We call our architecture state-denoised recurrent neural network (SD-RNN). We conduct a series of experiments to demonstrate the benefit of internal denoising, from small experiments like detecting parity of a binary sequence to larger natural language processing data sets. We characterize the behavior of the network using an information theoretic analysis, and we show that internal denoising causes the network to learn better on less data.