Separate global compaction from local eviction to keep KV-cache layouts stable under load (TokenPilot)
arXiv:2606.17016·low signal
TokenPilot (arXiv:2606.17016) splits context management into two layers — global ingestion-aware compaction and local lifecycle-aware eviction — to stabilize dynamic context layouts while conservatively offloading content based on task-level factors. The point is cache efficiency: naive eviction churns the KV cache and tanks throughput, whereas separating concerns keeps the cache prefix stable. Relevant if you're self-hosting agents at scale and watching prefill/cache-hit metrics.