Compute Globally, Materialize Locally: The Memory Contract of Sparse Event-KV

This paper tests a central assumption behind long-horizon agent memory: whether retained KV-cache entries remain informative after the observations that produced them are evicted. In controlled histories, some downstream cached rows follow an omitted observation even though the served context contains no explicit value identifying the correct answer. On Qwen3-8B, deliberately phrased answer-free events increased donor-aligned recovery from 6% to 51%. However, passively harvesting natural mentions from long-term dialogue showed no detected advantage. The authors also report bounded payload capacity and strong sensitivity to phrasing.