In this paper, we propose a new bricked cache system suitable for a particular surface propagation algorithm : seismic horizon reconstruction. The application domain of this algorithm is the interpretation of seismic volumes used, for instance, by petroleum companies for oil prospecting. To ensure the optimality of such surface extraction, the algorithm must access randomly into the data volume. This lack of data locality imposes that the volume resides entirely in the main memory to reach decent performances. In case of volumes larger than the memory, we show that using a classical brick cache strategy can also produce good performances until a certain size. As the size of these volumes increases very quickly, and can now reach more than 200GB, we demonstrate that the performances of the classical algorithm are dramatically reduced when processed on standard workstation with a limited size of memory (currently 8GB to 16GB). In order to handle such large volumes, we introduce a new slimming brick cache strategy where bricks size evolves according to processed data : at each step of the algorithm, processed data could be removed from the cache. This new brick format allows to have a larger number of brick loaded in memory. We further improve the releasing mechanism by filling in priority the “holes” that appear in the surface during the propagation process. With this new cache strategy, horizons can be extracted into volumes that are up to 75 times the size of the available cache memory. We discuss the performances and results of this new approach applied on both synthetic and real data.