The compliance by the general public to the nutritional guidelines issued by health authorities is relatively low and this represents a significant health concern. A solution to this problem could be to develop a nutritional recommender system. The challenge is to develop a system capable of suggesting not only foods that align with users' tastes but also menus, or even sequences of menus, that meet diverse restrictions such as particular diets, allergies, or calorie thresholds. We present a method based on recurrent neural networks that simultaneously model the sequential structure of meals and individual tastes. Our architecture is learned on the INCA2 dataset, which describes daily individuals' consumptions over a week. Our approach tackles the following research questions: (1) How meals are structured during the day? (2) How can we consider both sequential modelling and user preferences? and (3) How can we evaluate the performance of the recommendation in this context? Our experimental campaign employs several metrics to illustrate the successful incorporation of sequential aspects and user preferences in our models. This marks a crucial milestone in a larger project that aims to work with diverse input data and integrate various sequence constraints.