ShrikeSNT: A Temporal Stability Framework for Robust KV Cache Management in Large Language Models

The growing key-value (KV) cache in long-context autoregressive inference is a primary memory bottleneck. Existing compression methods treat the cache as a stateless queue, discarding tokens irreversibly without temporal continuity—making them unsafe under distribution shift. For example, H2O achieves only 1. 0% exact match on TriviaQA versus 51. 0% for the full-context baseline, and exhibits a +20 percentage point accuracy variance across 1k–16k token contexts. We introduce ShrikeSNT, a safe compression framework that reframes KV cache management as a stateful dynamical process. The novelty lies not in individual components but in their composition into a unified temporal stability framework that jointly implements: (i) memory inertia—EMA‑based importance accumulation to prevent eviction from transient noise; (ii) partial rollback—an explicit recovery mechanism that restores KV representations toward a recent snapshot; and (iii) phase‑gated activation—conditional suppression of compression when the cache is within normal bounds. In our experimental setting (greedy decoding, reproducible across independent runs), ShrikeSNT shows no measurable degradation within confidence bounds relative to the full‑context baseline on LongBench‑v2 (p = 1. 000, d = 0. 000, N = 300), while reducing time‑to‑first‑token by up to 2. 30 at the 14B scale. It is the only evaluated method that avoids catastrophic failure on short‑context tasks (TriviaQA: 50% vs. H2O 1%), recovers 75% of baseline ROUGE‑L on CNN/DailyMail, and maintains an accuracy profile identical to the uncompressed baseline across 1k–16k context lengths. ShrikeSNT establishes a safe compression regime—KV cache reduction without catastrophic degradation, performance variance, or positional bias.