Recurrent neural network models with parallel distributed architecture are constructed using ordinary differential equations. Models of associative memory are constructed using chaotic attractors. The learned memories are represented by arbitrary attractors, which can be oscillatory or chaotic.