A recent study introduces a novel training protocol for implicit neural representations designed to enhance scientific simulations by addressing the challenge of catastrophic forgetting. This approach leverages limited memory buffers alongside sketched data, enabling efficient in situ training without excessive memory consumption. The method is theoretically grounded in the Johnson-Lindenstrauss lemma, which supports the use of sketch-based regularization to maintain data integrity during continual learning. By integrating these elements, the protocol facilitates ongoing adaptation of neural compressors within scientific workflows. Its relevance extends to scenarios requiring continual learning, where maintaining performance over sequential data is critical. This development aligns with ongoing efforts to optimize neural network training under resource constraints, particularly in scientific domains. The approach represents a promising direction for improving the robustness and efficiency of implicit neural compressors in dynamic simulation environments.