iS-KV keeps recent key and value states exact while incrementally compressing older states into fixed-rank representations. Unlike eviction, every token position remains available if later reasoning attends to it again.
The key technique is updating historical coefficients whenever the shared low-rank basis changes, preventing old states from drifting into different reconstructions. On MATH-500, it reached 82.6% accuracy at 4.06-fold persistent-KV compression on DeepSeek-R1-Distill-Llama-8B, versus 83.6% uncompressed, and 89.2% at 5.64-fold compression on Qwen3-8B. It beat matched-memory R-KV and SnapKV at all six reported operating points.