Developing dedicated integrated circuits operating with low power consumption is indispensable to realize a large scale artificial neural network (ANN) like a human brain. Recently, emerging memories with non-volatile property have much attention for storing synaptic weights in ANNs. We have demonstrated associative memory operations using spin-orbit torque devices as non-volatile analog synaptic weights in Hopfield model. We confirmed that device mismatching can be compensated by learning process. Torrelable mismatching with respect to memory capacity was also evaluated by numerical simulation.