The memory problem is why I think post-rendered hair is preferable to geometry-based hair. If you take Eric's hair model, cut down the number of strands and increase their polygon count, maybe somehow you could calculate (prior to rendering) intermediate strands based on the weighted averages of the geometry strands. Then you could keep the actual geometry to a minimum, but the rendered hair would have many more strands than the geometry. Like if each geometry strand was the center of a hexagonal array of rendered strands, then you could have a ratio of rendered to geometry strands of 6:1, 18:1, 36:1, 60:1, 90:1 or even larger.