Cognitive Bifurcation: Dual-Progressive Causal Diffusion with Hippocampal Memory for Continual Graph Learning

Continual Graph Learning (CGL) on non-stationary streams faces the fundamental challenge of adapting to complex distribution shifts, where the entanglement of invariant causal structures and transient environmental noise inevitably leads to catastrophic forgetting. Under such non-stationary conditions, existing methods relying on incremental updates or raw replay are vulnerable to recursive error accumulation : a minor misinterpretation of the shift at an early stage propagates over time, causing a collapse in structural understanding. To tackle this, we draw inspiration from the cognitive bifurcation in the human brain and propose DCDHippo (Dual-Progressive Causal Diffusion with Hippocampal Memory). This framework treats adaptation as a closed-loop interplay between two systems. First, to handle real-time shifts, a Progressive Causal Masking mechanism (Fast System) dynamically prunes shift-induced noise to extract invariant causal skeletons. Simultaneously, to rectify local drifts, a Causal-Anisotropic Diffusion module (Slow System) internalizes these skeletons into a global invariant representation space via generative reconstruction. Crucially, we introduce the Evolving Hippocampal Memory that re-activates this global knowledge to distill wisdom back into the fast adapter, ensuring robust adaptation to continuous distribution shifts. Extensive experiments demonstrate that DCDHippo significantly outperforms state-of-the-art methods in both adaptability and knowledge retention.