Federated Learning (FL) enables collaborative model training without sharing raw data, but remains vulnerable to poisoning attacks from malicious clients. Existing defenses are often reactive and require costly model retraining, making them inefficient and impractical for real-time protection. We propose FedCleaner, a server-side dual-mechanism framework that combines: Proactive Layer-Wise Anomaly Detection to identify poisoned updates in real time; Retroactive Contribution Erasure to efficiently unlearn malicious client influences without retraining. Experiments on datasets show that FedCleaner provides a scalable, privacy-preserving, and regulation-compliant solution to defend FL systems against persistent poisoning threats.