Subjects in two experiments were asked to estimate how often items had appeared in a studied list. If the estimates are based on the value of a frequency attribute, the implicit question is, "How many times did you think of this item during study?" If, on the other hand, estimates are memorial attributions, the implicit question is, "How many study episodes does this item now make you think of?" In Experiment 1, subjects studied items that were repeated with the same or a different partner on each appearance. Items studied with a different partner each time received higher estimates than those with the same partner, but only if the items were semantic relatives of the partners. Thus, if tested items are strong cues for their partners, estimates of frequency benefit from contextual variety, implying that the partners are accessed during the test episode. In Experiment 2, several types of items were processed in several different ways. The result was that variables that enhance old-new discrimination also enhance the quality of estimates of frequency, implying that the two measures are based on the same cognitive stuff. Our conclusion is that judged frequency is attributed to items because of the memorial information they access at test.The research was supported by Natural Sciences and Engineering Research Council of Canada Grant A8122 to Ian Begg. We thank Douglas Hintzman, John Jonides, and Robert Rose for their detailed comments on an earlier version, and Larry Jacoby, Lee Brooks, and Andrea Snider for helpful discussions. We thank Larry Jacoby for suggesting the attribute versus attribution terminology as a way to capture the distinction we had in mind.