there is actually a paper which uses count min sketches to store very high dim means for k means clustering.