Towards a Combinatorial Characterization of Bounded-Memory Learning