This paper constructs a statistical model of learning that suggests a systematic way of measuring the persistence of treatment effects in education. This method is straightforward to implement, allows for comparisons across educational treatments, and can be related to intuitive benchmarks. We demonstrate the methodology using student-teacher linked administrative data for North Carolina to examine the persistence of teacher quality. We find that teacher-induced learning has low persistence, with three-quarters or more fading out within one year. Other measures of teacher quality produce similar or lower persistence estimates. (Contains 22 footnotes and 6 tables.)