PRISM is a training system for memory-augmented temporal graph neural networks that avoids memory staleness while still preserving strong GPU parallelism. It introduces multi-versioned memory refinement, a lightweight memory computation graph, and a high-throughput temporal sampling pipeline for dynamic graphs.
Memory-augmented temporal graph neural networks are powerful for dynamic graphs, but their training becomes difficult when events inside the same batch depend on one another. Existing high-throughput systems commonly ignore those intra-batch dependencies, which creates stale memory and harms accuracy as batch size grows.
PRISM addresses this by computing multiple fresh memory versions inside a batch so each event can consume a temporally consistent state. The result is a training pipeline that preserves model semantics more faithfully while still scaling well on GPUs.
Standard stale-memory training improves throughput by processing events together, but it can violate temporal consistency. PRISM shows that accuracy losses at large batch sizes are not inevitable; they are often a systems problem caused by stale memory.
Across five temporal-graph benchmarks and three unmodified models, PRISM is reported to improve accuracy while keeping training time competitive with parallel stale-memory baselines and often lower than stricter staleness-free methods.
Training memory-augmented temporal graph neural networks efficiently and accurately remains challenging because memory staleness appears when temporally dependent events are processed together in the same batch. PRISM avoids that staleness by using multi-versioned memory vectors, allowing each event to read a temporally consistent memory state. The system formalizes lazy freshness, implements a parallel memory refinement algorithm, and combines it with efficient temporal sampling support for dynamic graphs. The paper reports strong accuracy gains while maintaining competitive training efficiency.