This study quantified human short-term-memory decay functions for delayed vibrotactile frequency discriminations. Subjects indicated which of two successive intervals contained the higher or lower frequency of a pair separated by delay periods of 0.5-30 sec. Performance decreased as a function of length of delay and was higher when delays were unfilled than when they were filled with a backwards-counting task. This interpolated task may have interfered with rehearsal of a coded representation of the remembered vibrotactile frequency. A change in decay rate after 5-sec delays suggests a switch from reliance on sensory memory to the coded frequency representation. Performance and decay rate depended on presentation order of higher or lower frequency within pairs. Reciprocal performance asymmetries seen in high- versus low-frequency ranges did not result from simple response bias.