The formation of behavioral traditions has been considered as one of the main building blocks of culture. Numerous studies on social learning in different animal species provide evidence for their capability of successful transmission of information. However, questions concerning the memory and maintenance of this information have received comparably little attention. After the innovation and initial spread of a novel behavior, the behavior should stabilize and be maintained over time. Otherwise, the behavioral pattern might collapse and no tradition formation would be possible. The aim of this study was to investigate long-term preferences in a two-action manipulation task in common marmosets (Callithrix jacchus). Three captive family groups (23 individuals in total) were trained on one of two possible techniques to open a wooden box and gain access to a food reward, by either pulling or pushing a flap door. The training phase took place in a family group setting, while the test phase was conducted individually. Although the subjects could experience the alternative technique during the test sessions, the majority preferentially used the technique learned in the group setting. Moreover, the subjects were re-tested six times over a period of more than four years, in order to examine the fidelity of their preferences. The longest break without exposure the task lasted for 3.5 years. In all tests, the marmosets showed a similar preference as in the first test block shortly after the training. To our knowledge, this is the first lab study that experimentally demonstrates memory and fidelity of experimentally seeded information in a manipulation task over a time period of several years, supporting the assumption that socially learned foraging techniques can lead to relatively stable behavioral traditions.