This research investigated the kind of semantic information that facilitates the relearning of a nonrecalled item (i.e., the kind of semantic information that is saved in the memory trace of a forgotten item). The paradigm consisted of three stages: (a) learn a list of number-word pairs (e.g., 95-donate), (b) return approximately 1 month later for a retention test, and (c) relearn a new list of numberword pairs that have various kinds of semantic relatedness to the originally learned pairs (e.g., 95-donate for an identity relationship, 95-give for a relationship along some specific dimension such as synonymy, or 95-profess for an unrelated control relationship). Considering relearning performance on only those items not recalled on the retention test (i.e., forgotten items), a relearning advantage of the identity condition over the control condition demonstrates that some kind of information was saved; a relearning advantage of the specific-dimension condition (e.g., synonymy) over the control condition demonstrates that the specific information under investigation (e.g., synonymic information) was saved in the memory traces of forgotten items.Seven experiments investigated various kinds of semantic information that potentially could be saved in a nonrecalled memory trace. Experiment 1 showed a lack of reliable savings for both synonymic information (e.g., despises-hates) and antonymic information (e.g., loves-hates). Experiment 2 yielded reliable savings for both superordinate information (e.g., buick~car) and subordinate information (e.g., vehicle-car). Experiment 3 suggested that information is saved about some (i.e., more than one, but not all) subordinates of the originally learned item. Experiment 4 extended the lack of synonymic savings to three parts of speech: verbs, nouns, and adjectives. Experiment 5 showed that the lack of synonymic savings occurs not only when acquisition was via verbatim recall, but also when acquisition was via gist recall. Experiment 6 extended this finding by demonstrating a lack of synonymic savings-even when the learner, studying for gist recall, was forced to generate a synonym of each item on every acquisition study trial! Experiment 7 yielded a lack of reliable savings for associative information (e.g., table-chair).A parsimonious theoretical account of the above results for five major kinds of semantic information is given by a proposition that derives from the hierarchical notion of inclusion (e.g., buicks are included in cars, which, in turn, are included in vehicles; cars have approximately the same degree of inclusion as automobiles and as trucks). This proposition is: Only semantic information at greater and lesser degrees of inclusion (than the originally learned item) is saved in the memory trace during forgetting. That is, superordinate information and subordinate information are saved (greater and lesser inclusion, respectively), whereas synonymic informa-
