In this paper, a new type of local search algorithm is proposed, called Learning Tabu Search and denoted LTS. It is assumed that any solution of the considered problem can be represented with a list of characteristics. LTS involves a learning process relying on a trail system. The trail system is based on the idea that if some combinations of characteristics often belong to good solutions during the search process, such combinations of characteristics should be favored when generating new solutions. It will be showed that LTS obtained promising results on a refueling problem in a railway network.