The Alignment Tax on Continual Learning: Inverse Scaling of Memory Consolidation in Language Models | Synapse