Another potential weakness that isn't immediately clear from this experiment is if the experiment was run much longer (disregarding cost) then perhaps then the agent's memory could be susceptible to more long term memory compaction corruption and thus made more compliant?