Passivity is one of key characterizations of the Lagrange equation of motion for various mechanical systems including robot arms. After introducing a set of postulates for learning control, this paper points out the importance of passivity in learning control for refinement of motions in mechanical systems. In particular it is shown that, given a desired output motion trajectory, the exponential passivity with a quadratic margin for residual error dynamics between the actual motion and the desired one plays a crucial role in steady improvement of motions. Hence it is claimed that this properly deserves to be called“learnability”(inherent ability of learning) because it implies the convergence of motion trajectories to the desired one. In addition, robustness problems of learning are discussed when the existence of initialization errors and fluctuations of dynamics are permitted to some extent. It is shown that introduction of a forgetting factor into the update law of learning is crucial in assuring the uniform boundedness of motion trajectories and the convergence of them in an ε-neighborhood of the desired one. Finally a special class of learning methodology called“selective”learning which updates the content of a long-term memory every after several trials is proposed in order to accelerate the speed of convergence.