Although catchment behaviour during recession periods appears to be better identifiable than in other periods, the representation of hydrograph recession is often weak in hydrological simulations. Reason lies in the various sources of uncertainty that affect hydrological simulations, and in particular in the inherent uncertainty concerning model conceptualizations, when they are based on an a-priori representation of the natural system. When flawed conceptualizations combine with calibration strategies that favour an accurate representation of peak flows, model structural inadequacies manifest themselves in a biased representation of other aspects of the simulation, such as flow recession and low flows. In this paper we try to reach good model performance in low flow simulation and make use of a flexible model structure that can adapt to match the observed discharge behaviour during recession periods. Moreover, we adopt a step-wise calibration procedure where we try to avoid that the simulation of low flows is neglected in favour of other hydrograph characteristics. The model used is designed to reproduce specific hydrograph characteristics and is composed of four reservoirs: an interception reservoir, an unsaturated soil reservoir, a fast reacting reservoir, and a slow reacting reservoir. The slow reacting reservoir conceptualises the processes that lead to the generation of the slow hydrograph component, and is characterized by a storage-discharge relation that is not determined a-priori, but is derived from the observations following a ``top-down'' approach. The procedure used to determine this relation starts by calculating a synthetic master recession curve that represents the long-term recession of the catchment. Next, a calibration procedure follows to force the outflow from the slow reacting reservoir to match the master recession curve. Low flows and high flows related parameters are calibrated in separate stages because we consider them to be related to different processes, which can be identified separately. This way we avoid that the simulation of low discharges is neglected in favour of a higher performance in simulating peak discharges. We have applied this analysis to several catchments in Luxembourg, and in each case we have determined which form (linear or non linear) of the storage-discharge relationship best describes the slow reacting reservoir. We conclude that in all catchments except one (where human interference is high) a linear relation applies.