The authors studied longitudinal change in learning efficiency as a predictor of future dementia type among healthy, well-educated, noninstitutionalized elderly retirees. Serial assessments of memory were obtained using the California Verbal Learning Test (CVLT). Latent growth (LG) models were developed from the slopes of the subjects' performance over the first five CVLT learning trials at each of three serial administrations (e.g., cohort inception [i.e., baseline] [CVLT1], 18 months [CVLT2] and 36 months [CVLT3]). The resulting growth curves were incorporated into a higher order LG model representing the dynamic change in learning efficiency over time (DeltaCVLT). DeltaCVLT was used to predict each subject's "dementia type" (i.e., clinical state) at 36 months (e.g., no dementia, Type 1 [Alzheimer type] dementia or Type 2 [non-Alzheimer type] dementia), after adjusting for CVLT1, baseline age, and baseline dementia type. Nonlinear (logarithmic) LG models of CVLT1-CVLT3 and DeltaCVLT best fit the data. There was significant variability about both CVLT1 and DeltaCVLT, suggesting subgroups in the sample with significantly different baseline memory function, and different rates of deterioration in learning efficiency. Age, baseline dementia type, and DeltaCVLT made significant independent contributions to final dementia type. CVLT1 did not predict final dementia type independently of the other covariates. These data suggest that baseline memory performance in noninstitutionalized elderly retirees does not predict future dementia type independently of the dynamic rate of change in memory measures. Serial administrations of memory tests may help identify nondemented persons at greater or lesser risk for conversion to frank dementia in the near-term.