When making memory predictions (judgments of learning; JOLs), people typically underestimate the recall gain across multiple study–test cycles, termed the underconfidence-with-practice (UWP) effect. This is usually studied with verbal materials, but little is known about how people repeatedly learn and monitor their own actions and to what extent retrieval practice via interim tests influence the progression of JOLs across cycles. Using action phrases (i.e., squeeze the lemon) as learning material, we demonstrated the UWP effect after both verbal and enactive encoding, although we did not get first-cycle overconfidence. As predicted, participants exhibited underconfidence in Cycles 2 and 3, as an error of calibrations. However, people’s resolution of JOLs (i.e., ability to discriminate recalled from unrecalled items) increased across study–test cycles. Importantly, JOLs for study–test (relative to study–study) items increased faster across cycles suggesting that repeated study–test practice not only produces underconfidence across cycles, but also reduces underconfidence relative to study–study practice. We discuss these findings in terms of current explanations of the underconfidence-with-practice effect.