• E-CGL introduces a combined replay sampling strategy based on node importance and diversity to mitigate catastrophic forgetting in continual graph learning.
• By sharing weights between a GNN and a lightweight MLP, E-CGL bypasses expensive message-passing, yielding average training and inference speedups of 15.83× and 4.89×.
• The method achieves state-of-the-art results on four CGL datasets, reducing average catastrophic forgetting to −1.1%.
• E-CGL effectively handles topological interdependencies between sequential graph snapshots while maintaining efficiency at scale.