By allowing models to actively update their weights during inference, Test-Time Training (TTT) creates a "compressed memory" that solves the latency bottleneck of long-document analysis.