Replicated systems are a kind of distributed systems whose main goal is to
ensure that computer systems are highly available, fault tolerant and provide
high performance.  One of the last trends in replication techniques managed by
replication protocols, make use of Group Communication System, and more
specifically of the communication primitive atomic broadcast for developing
more efficient replication protocols.

An important aspect in these systems consists in how they manage the
disconnection of nodes --which degrades their service-- and the
connection/reconnection of nodes for maintaining their original support. This
task is delegated in replicated systems to recovery protocols. How it works
depends specially on the failure model adopted. A model commonly used for
systems managing large state is the crash-recovery with partial amnesia because
it  implies short recovery periods. But, assuming it implies arising several
problems. Most of them have been already solved in the literature: view
management, abort of local transactions started in crashed nodes --when
referring to transactional environments-- or for example the reinclusion of new
nodes to the replicated system. Anyway, there is one problem related to the
assumption of this second failure model that has not been completely
considered: the amnesia phenomenon. Phenomenon that can lead to inconsistencies
if it is not correctly managed.

This work presents this inconsistency problem due to the amnesia and formalizes
it, defining the properties that must be fulfilled for avoiding it and defining
possible solutions.  Besides, it also presents and formalizes an inconsistency
problem --due to the amnesia-- which appears under a specific sequence of
events allowed by the majority partition progress condition that will imply to
stop the system, proposing the properties for overcoming it and proposing
different solutions. As a consequence it proposes a new majority partition
progress condition.  In the sequel there is defined in a more accurate and
precise way one of the solutions -- that consists in persisting messages
atomically in the delivery process-- and it is studied its behaviour under
different replicated systems configurations.  Finally, it is included a study
of the overhead introduced  when using this generic solution, demonstrating
that if the memory is fast enough the overhead can be tolerated.