Distributed data management services usually run on top of dynamic and heterogeneous systems. This remains true for most distributed services. At large scale, it becomes impossible to get an accurate global view of the system. To provide the best quality of service despite this highly dynamic environment, those services must continuously adapt. To do so, they monitor their environment and store events and states of the system in memory. In this paper, we propose a model to formalize this memory and three strategies to use it. We explain the theoretical difference between those strategies and conduct an experimental evaluation. We show that providing the ability to "forget" old events leads to better performance. However, using fading events (events that progressively disappear) rather than events that suddenly disappear leads to even better performance and is more adequate to detect habits and recurrent behaviors.