That would be the naive approach which, at first, sounds good but doesn't bode all that well from a hardware standpoint. Since a frame of animation is in reference to the previous frame it is possible to store the transition information in a transformation matrix and use that as a keyframe. If you were to store each position as a mesh then you have tons of wasted space because perhaps not all vertices are affected and yet you reserve space for every one of them regardless of their part in the frame. This also means you have to push these vertices across the bus and on complex scenes with many characters this would slow to a halt.
The book im reading right now suggests that you use an array of mesh's just as you would with sprite animation and index into the array to choose which mesh to render.