ChurnBench: in agentic RAG, the TTL-to-change-rate ratio drives staleness, not cache age
arXiv 2609.11515·medium signal
The benchmark generates a four-source enterprise data fabric as a timeline with an append-only ground-truth ledger, so it can tell freshness errors apart from reasoning errors. With scheduled refresh, freshness errors were 7, 4 and 4 at cache ages of 1, 14 and 28 days. Disabling tiered refresh raised 28-day errors from 4 to 45. Tune each entity's TTL to how often that entity changes, such as licenses, prices or users, rather than setting one global cache age.