The cost and Achilles' heel of long-context LLM inference both live in the KV cache — the moment compression ratio is pushed up, multi-hop reasoning collapses first. PReM (Preserve and Refresh Memory, arXiv:2607.14327) from Tongyi Lab's Zheng Bo and others gives a clean answer: instead of pursuing one-shot static compression, treat the long context as a layer-wise KV memory inside the model and refresh it on demand during generation. PReM has three components: a dedicated "memory layer" inserted in the middle of the Transformer scores chunks in real time, keeping only the evidence truly needed by the current step; a special token <m> is introduced — once the model outputs it, it triggers a cross-layer KV-memory re-selection; the Top-k chunks retain their original KV, the rest are mean-pooled into a single representative vector (Preserve-and-Pool), balancing detail and redundancy under a fixed budget. The training side has a paired "phase-separated refresh training", slicing inference into memory-selection and conditional-generation stages, and using contrast loss and boundary loss to force the model to identify evidence and guarantee generation consistency across refreshes. On 32K contexts, PReM simultaneously beats SnapKV, CAKE, LongLLMLingua, EXIT and other baselines at 16× and 32× compression ratios; the gain on multi-hop Q&A is particularly striking, with a 3B small model overtaking larger solutions. The signal worth remembering: dynamic on-demand refresh may be closer to the real bottleneck of long-context inference than the compression-ratio race.