Memory module improves efficiency of long-context AI models

Original: Augmenting Attention with Exponentially Decaying Memory Improves Query-Aware KV Sparsity

Source: arxiv.org ↗