Children with higher working memory or updating (WMU) capacity perform better in math. What is less clear is whether and how this relation varies with grade. Children (N = 673, kindergarten to Grade 9) participated in a 4-year cross-sequential study. Data from 3 WMU (Listening Recall, Mr. X, and an updating task) and a standardized math task (Numerical Operations) showed strong cross-sectional correlations at each of the 10 grades, but particularly at Grades 1 and 2. Cross-lagged autoregressive analysis showed invariance in the predictive relations between WMU and subsequent math performance, but the importance of domain-specific knowledge increased with grade. Latent growth modeling showed that higher WMU capacity at kindergarten predicted higher math growth rates, averaged across all grades, but WMU growth rate was invariant across grades. Socioeconomic status, but not gender, explained variance in WMU at kindergarten. Implications for WM training are discussed.