Subjects learned one of four patterns of like-dislike relations between members of four-person structures. They then recalled the learned structure after an interval of fifteen minutes, one week, or four weeks. One pattern of relations was balanced, one could be balanced by changing only one pair of relations, and the other two were totally unbalanced. There was clear evidence that the balance assumptions were employed in learning; for unbalanced structures, there was some evidence of the use of positivity and of source/target generalization. Subjects recalled the balanced structure more accurately than the unbalanced ones, especially over the longest time interval. However, unbalanced structures were not converted in memory to balanced ones.