KV‑cache reuse can cut inference cost in retrieval‑augmented generation (RAG) and agentic systems, but when retrieved knowledge, working memory, or user state is edited, the cached context becomes stale. Because causal self‑attention propagates even a local edit downstream, a full re‑prefill restores consistency at high cost, while refreshing only the edited span may leave downstream dependencies outdated.
We cast in‑place repair as a budgeted recomputation problem and evaluate several training‑free position‑selection policies on a factual RAG benchmark that contains matched direct and derived edits. Across three model families, every policy fixes direct edits, but derived edits clearly separate their performance. With the primary budget, a contiguous edit‑local window recovers at least 0.94 of the post‑edit answer margin and substantially outperforms attention‑based, KV‑deviation, and structural selectors.
Mechanistic analysis shows that position sets effective under clean‑state transplantation can fail during actual recomputation because scattered positions inherit surrounding staleness. The edit‑local advantage also depends on adjacency and largely disappears when the answer‑bearing text moves downstream.
Since answer‑relevant edits almost always corrupt model behavior, their failure severity is hard to predict, and repair is 13‑21× faster than full re‑prefill, our findings support unconditional edit‑local repair when the dependent text remains adjacent to the edit.
Review