In the past, several attempts have been made to deal with the state space explosion problem by equipping a depth-first search (DFS) algorithm with a state cache, or by avoiding collision detection, thereby keeping the state hash table at a fixed size. Most of these attempts are tailored specifically for DFS, and are often not guaranteed to terminate and/or to exhaustively visit all the states. In this paper, we propose a general framework of hierarchical caches which can also be used by breadth-first searches (BFS). Our method, based on an adequate sampling of BFS levels during the traversal, guarantees that the BFS terminates and traverses all transitions of the state space. We define several (static or adaptive) configurations of hierarchical caches and we study experimentally their effectiveness on benchmark examples of state spaces and on several communication protocols, using a generic implementation of the cache framework that we developed within the CADP toolbox.