Recommendation systems typically require continual adaptation to take into account new and obsolete items. A major challenge in this situation is catastrophic forgetting, where the trained model forgets patterns it has learned before. We propose a method to mitigate this effect.