Resistive random-access memories (RRAM)s are considered a promising candidate for neuromorphic circuits and systems. In this letter, we investigate using TiO-2 RRAMs to solve blind source separation problem through independent component analysis (ICA) for the first time. ICA has numerous uses including feature extraction. We deploy a local, unsupervised learning algorithm (error-gated Hebbian rule) to extract the independent components. The online evaluation of the weights during the training is studied taking into consideration the asymmetric nonlinear weight update behavior. The effects of the device variability are considered in the results. Finally, an example of de-mixing two Laplacian signals is given to demonstrate the efficacy of the approach.