<p>With the rise of passive sensing technology, significant emphasis is placed on preserving the privacy of human activity recognition (HAR) while ensuring the extensibility of action analysis. Most HAR methods require the same action categories to be used in both training and testing, thereby limiting the extensibility of these systems. To address this limitation, this paper proposes an extensible HAR system that enables users to dynamically add or remove specific action categories. To the best of our knowledge, this is the first work to achieve extensible activity recognition by allowing the addition and removal of actions while operating within the constraints of predefined action categories in the field of WiFi sensing. Wiear introduces a class-incremental training and update strategy during the addition of new actions to mitigate catastrophic forgetting of previously learned action categories. Furthermore, a machine unlearning-based training and update strategy is employed during the removal of existing actions, enabling the selective forgetting of specific actions while preserving the integrity of retained categories. Additionally, this study conducts extensive comparative and case studies to evaluate Wiear against other systems in fixed action category scenarios. Experimental results demonstrate that Wiear maintains an average accuracy of 96.62% despite multiple class additions and removals, while significantly reducing computational overhead in terms of execution time and storage requirements. These advantages enhance its suitability for real-world applications requiring extensible user action recognition.</p>