In machine learning, the ability for models to selectively forget specific data is increasingly crucial. Traditional “machine unlearning” methods often involve extensive retraining, which is computationally expensive and impractical for real-time applications. This paper introduces Federated Meta Unlearning (FMU), a novel approach within the Federated Learning (FL) framework that merges the adaptability of meta-learning with the precision of machine unlearning. FMU utilizes a modified Model-Agnostic Meta-Learning algorithm, employing model decomposition and weighted aggregation techniques. This approach enables efficient data influence removal while minimizing impact on overall model performance. The key innovation of FMU is the decomposition of the federated model into adjustable components, allowing targeted modifications or removal based on unlearning requests, followed by weighted re-aggregation to maintain model integrity and flexibility. Extensive experiments across diverse datasets show that FMU reduces the computational burden of traditional unlearning while preserving high model performance and adhering to privacy standards. Additionally, FMU enhances adaptability to new data and tasks, making it suitable for dynamic and privacy-sensitive settings. In summary, FMU offers a substantial advancement in addressing the challenge of selective forgetting in machine learning.