As trust management can address the new security challenges in pervasive computing environments, we present our pervasive trust model TM-RMRP which is composed of input factors and trust evaluation. With the model presented, anonymous entities can collaborate with trustful parties without central security managers. To reduce or even avoid inappropriate or inaccurate recommendations, we introduce the concept of recommendation trust and propose a recommendation mechanism of rewards and punishments based on this concept to identify more reliable recommendations. From the simulation results, the mechanism is proved to be applicable in pervasive computing environments.