A method to embed P multi-valued patterns x ^s  ∈ ℝ ^N into a memory of a recurrent neural network is introduced. The method represents memory as a nonlinear line of attraction as opposed to the conventional model that stores memory in attractive fixed points at discrete locations in the state space. The activation function of the network is defined by the statistical characteristics of the training data. The stability of the proposed nonlinear line attractor network is investigated by mathematical analysis and extensive computer simulation. The performance of the network is benchmarked by reconstructing noisy gray-scale images.