Presents a postscript to the current authors' comment on the original article, "Short-term memory for serial order: A recurrent neural network model," by M. M. Botvinick and D. C. Plaut. In their commentary, the current authors demonstrated that Botvinick and Plaut's (2006) model of immediate serial recall catastrophically fails when familiar letters are tested in untrained positions within a list (Simulation 2), and a modified version of their model with a distributed letter coding scheme also fails to recall familiar (and novel) letters when tested in untrained positions (Simulation 7). That is, short-term memory (STM) did not generalize to all possible test sequences. Botvinick and Plaut rejected these claims and reported a simulation in which a new version of their model recalls familiar (and novel) items in novel positions. What Botvinick and Plaut did not emphasize, however, was that the model was trained on all the letters in all positions of the list. So, in principle, the model could recall novel syllables (and familiar syllables in untrained positions) by recalling familiar phonemes in trained positions. The current authors ran two new simulations to further highlight the generalization constraints associated with these learned conjunctive codes. (Contains 2 figures.)