<p>Variations in deployment parameters such as new users or sensor locations are a major concern in sensor-based human activity recognition (HAR), accounting for a major source of performance drop from in-lab experiments to in-the-wild deployment. While fine-tuning a pre-trained model using data acquired during deployment can reduce such performance loss, it often leads to catastrophic forgetting. The hardware constraints of edge devices also limit the options of continual learning techniques. To address these challenges, we introduce COOL, a continual online on-device learning method leveraging Kolmogorov-Arnold Networks (KANs). Our method exploits the inherent plasticity of KANs. COOL is evaluated through two HAR scenarios (utilizing bio-impedance and Inertial Measurement Unit (IMU) signals separately) to demonstrate its performance in addressing both the catastrophic forgetting issue and concept drift issue caused by new targets (users and sensor locations). A significant average overall performance improvement of around 6.82% and 9.80% in upper- and lower-body fitness activity recognition and 16.65% improvement in Locomotion activity recognition was observed from the experiment compared to the SotA methods. Specifically, the on-device trained model demonstrated the memory-retention ability with a forgetting measurement score of less than 0.01, 100-fold better than SotA on-device HAR continual learning methods. Finally, we evaluate the hardware performance of the proposed COOL model with the STM32H743 microcontroller, an on-device inference and training with a time cost of 44.30 ms and 46.03 ms is observed, guaranteeing real-time model adaptation.</p>