Large-scale distributed systems are highly dynamic: nodes may crash, messages may be delayed or lost, new nodes may join, virtual machines may migrate... The workload may also vary a lot depending on the users or applications behavior. Thus, distributed systems and services must adapt continuously in order to remain efficient. To do so, distributed systems usually monitor their environment and maintain a knowledge of what has happened (e.g., heartbeats time-stamps for a failure detection mechanism, data access statistics for a distributed storage systems, etc.). Based on these observations, a distributed system can decide to adapt to better tackle future situations (migrate a virtual machine, create a new data replica, etc.). However, even if a system has a full event log, it is not a trivial task to decide how much and which part of the past must be taken into account. In this paper we propose and study the impact of multiple approaches, from "full-memory" to "no-memory" through "time-window" and "fading-memory" based strategies.