Objective: Sleep disturbance is common in dementia, although it is unclear whether differences in sleep architecture precede dementia onset. We examined the associations between sleep architecture and the prospective risk of incident dementia in the community-based Framingham Heart Study (FHS). Methods: Our sample comprised a subset of 321 FHS Offspring participants who participated in the Sleep Heart Health Study between 1995 and 1998 and who were aged over 60 years at the time of sleep assessment (mean age 67 ± 5 years, 50% male). Stages of sleep were quantified using home-based polysomnography. Participants were followed for a maximum of 19 years for incident dementia (mean follow-up 12 ± 5 years). Results: We observed 32 cases of incident dementia; 24 were consistent with Alzheimer disease dementia. After adjustments for age and sex, lower REM sleep percentage and longer REM sleep latency were both associated with a higher risk of incident dementia. Each percentage reduction in REM sleep was associated with approximately a 9% increase in the risk of incident dementia (hazard ratio 0.91; 95% confidence interval 0.86, 0.97). The magnitude of association between REM sleep percentage and dementia was similar following adjustments for multiple covariates including vascular risk factors, depressive symptoms, and medication use, following exclusions for persons with mild cognitive impairment at baseline and following exclusions for early converters to dementia. Stages of non-REM sleep were not associated with dementia risk. Conclusions: Despite contemporary interest in slow-wave sleep and dementia pathology, our findings implicate REM sleep mechanisms as predictors of clinical dementia.