<p>Money laundering involves complex, multi-account transaction patterns that often span multiple financial institutions. While cross-bank collaboration is crucial for effective Anti-Money Laundering (AML), privacy regulations prohibit direct data sharing, limiting traditional centralized solutions. Furthermore, evolving laundering behaviors cause static Graph Convolutional Network (GCN)-based models to suffer from catastrophic forgetting in temporal AML settings. Existing continual graph learning approaches alleviate forgetting but typically rely on memory buffers, raising scalability and privacy concerns in real-world financial settings. To address these challenges, we propose FC2L, a hybrid Federated Continual Contrastive Learning framework for graph-based AML that integrates GCN-based representation learning with federated optimization, enabling cross-institutional edge-level AML detection without sharing raw sensitive data. FC2L introduces a memory-free Column-wise Adaptive Balanced Orthogonal Projection (ACOP) mechanism that exploits direction-wise feature correlations to preserve task-relevant subspaces from past AML tasks while adapting to emerging patterns. To mitigate data heterogeneity across banks, we further design a dual task-specific contrastive learning objective that jointly aligns edge embeddings and classifier parameters in a shared feature space, enhancing representation consistency during continual learning. We theoretically show that ACOP achieves a favorable stability-plasticity trade-off compared to existing methods. Extensive experiments on benchmark AML datasets demonstrate that FC2L consistently outperforms both federated and continual learning baselines in representation quality, knowledge retention, and classification performance.</p>