The correct functioning of a peer-to-peer network relies on cooperative behavior of peers as service providers. Current approaches to detection and deterrence of non-cooperative behavior, such as reputation systems, rely on (1) global sharing of observations about service provisioning and (2) global authentication. These two factors severely impair the practical applicability. We propose a novel forward feedback protocol that is completely local and authentication-free and where peers locally and independently learn to avoid non-cooperative peers and routing failures. We evaluate our system in a variety of failure scenarios. The convergence rates and failure resilience are close to those of a fully centralized reputation system. High churn is tolerated without significant drop in performance. For increased fault tolerance the system takes advantage of the service replicas existing in the network. The proposed protocol is lightweight and can readily be integrated into any architecture where service requests are recursively routed, which includes all modern structured overlays.