Carryover: The Reuse Margin of a Stale KV Cache in Long-Horizon Agents
First author · NeurIPS 2026 Workshop on Long-Context Foundation Models
When a long-horizon agent hands off, can the next step reuse the cache the last one left behind? We measured it at 35 recorded coding-agent handoffs with 35K–80K-token histories. At a tolerance fixed before any run, the median same-model removal fraction is zero — replicated across three models — yet reused states still shift next-token predictions (median 90.2% top-1 agreement with a fresh cache). Cache agreement and prediction sensitivity have to be checked separately; meeting one does not guarantee the other. The study was pre-registered: hypotheses and decision rules committed before any run, results on a hash-chained ledger, outcomes reported as decided, including the ones the data would not support.