Deviations from model-based predictions of strategies leading to stable cooperation between unrelated individuals have raised considerable debate in regards to decision-making processes in humans. Here, we present data on cleaner wrasse (Labroides dimidiatus) that emphasize the importance of generalizing this discussion to other species, with the aim to develop a coherent theoretical framework. Cleaners eat ectoparasites and mucus off client fishes and vary their service quality based on a clients’ strategic behaviour. Hitherto, cognitive tasks designed to replicate such behaviour have revealed a strong link between cooperative behaviour and game theoretic predictions. However, we show that individuals from a specific location within our study site repeatedly failed to conform to the published evidence. We started exploring potential functional and mechanistic causes for this unexpected result, focusing on client composition, cleaner standard personality measures and ontogeny. We found that failing individuals lived in a socially simple environment. Decision rules of these cleaners ignored existing information in their environment (‘bounded rationality’), in contrast to cleaners living in a socially complex area. With respect to potential mechanisms, we found no correlations between differences in performance and differences in aggressiveness or boldness, in contrast to results on other cooperative species. Furthermore, juveniles from the two habitat types performed similarly, and better than the adults from the socially simple environment. We propose that variation in the costs and benefits of knowledge may affect a cleaners’ information acquisition and storage, which may explain our observed variation in cooperation and cognition.