Transactional replicated systems have usually adopted the crash-recovery with partial amnesia failure model. These systems need to deal with the amnesia phenomenon; i.e. non committed state is lost at crash time. So, if this phenomenon is not accurately managed in recovery processes it can lead to state inconsistencies in the replicate state. A general solution that consists in persisting messages atomically in the delivery process has been proposed for overcoming this problem, demonstrating also its validity. But, its use implies a cost: the overhead introduced for persisting messages atomically in the delivery process. This paper analyses this overhead simulating the proposed solution for a transactional replication protocol based on certification and demonstrates how this overhead can be minimized using solid-state memories, making it acceptable.