This report proposes an on-line compression algorithm for distributed executions. An execution is decomposed into atomic communication patterns. Events are clustered together according to the following policy: the initial partial order structure is preserved, and valid global states are detected in order to find a coherent global state if a site fails. Two versions of the algorithm have been developed. The first version is online and centralized and does not depend on execution order of any concurrent events. Moreover, if clusters are bounded in size, the memory size needed for infinite executions is also bounded. The second version is distributed and is done by piggybacking subset of sites knowledge into messages: under certain conditions, size of piggybacked messages is bounded and the algorithm can be effectively implemented.