Digital reminder systems are widely used to support daily routines among older adults, especially people living with dementia (PLwD). While effective at issuing prompts, these systems typically lack mechanisms to monitor user engagement or detect subtle behavioral changes that may indicate early signs of dementia progression or need for caregiver intervention. This study presents a proof-of-concept for transforming digital reminder systems into behavioral anomaly detectors using LSTM (Long Short-Term Memory) autoencoders. Through simulated datasets generated from smart home interaction logs, we model two caregiver-prioritized anomalies—delayed acknowledgments and location-based ignoring of reminders. We evaluate the performance of vanilla LSTM autoencoders, classifier-augmented variants, and traditional methods such as Z-score thresholding and Isolation Forest. Our results show that the LSTM autoencoder combined with a Random Forest classifier consistently outperforms all other models across all anomaly types and severity levels, achieving the highest detection performance in every condition tested. This work demonstrates the feasibility of using existing reminder systems as privacy-preserving, low-friction platforms for early behavioral monitoring in dementia care, with implications for enhancing caregiver support and clinical insight without requiring additional hardware or invasive sensing.