Attic-KV Cache Compression via Self-Testing Over Rehearsal
October 9, 2026
Rereading entire contexts to select KV cache entries fails under tight budgets, with a 3% keep ratio yielding only 31.5/96.5 on RULER. The authors suggest that caches improve memory more effectively through self-testing mechanisms rather than simple rehearsal-based compression.
HOW THIS AFFECTS YOU
●
builderYou should reconsider standard rehearsal-based KV cache compression strategies for long-context retrieval.
●
researcherThis challenges the prevailing assumption that more complete context rehearsal leads to better memory retention.