Cognitive impairment is often diagnosed only after substantial neurodegeneration. Overnight polysomnography (PSG) records brain, cardiac, respiratory and muscular activity, and may contain signals related to later cognitive decline. The 2026 George B. Moody PhysioNet Challenge focused on predicting future cognitive impairment from clinical sleep studies. A night of PSG is a long multimodal sequence; its segments are not equally informative, labels are scarce, and channel configurations vary between sites. We developed a complete prediction pipeline that used a frozen SleepFM encoder for five-minute PSG segments, a bidirectional LSTM for temporal modelling, and annotation-guided attention based on algorithmic sleep stages and events. Demographic variables and whole-night summary statistics were also fused for the final prediction. In the held-out local evaluation, the final model achieved an Age-Conditioned AUROC of 0.715 on I0002. Our team, Everybody Hearts (REM), obtained a Challenge score of 0.705 on the hidden site I0004 and XXX on the hidden site I0007 (rank XXX of XXX), demonstrating the potential of this foundation-model-based pipeline for future cognitive-impairment prediction.