Analysis
AKV, a research system from Tsinghua University and the Chinese University of Hong Kong, gives agents control over which key-value cache entries remain active. Its approach preserves retained computed states rather than summarizing text and recomputing the history that follows.
The researchers report 27.7% lower average inference cost and a 1.7-point increase in average benchmark score across three models and four benchmarks. These are study results. The cost comparison uses API-equivalent estimates; it is not a measured reduction on every provider's invoice.
Our reading: memory policy is becoming part of agent design. The question is what the agent can safely stop attending to while retaining enough information to finish its task. The paper was submitted October 20 and surfaced in today's clipping.
Source: Cache Engineering for Agents — AKV paper and project · 2026-10-20
Why it matters
Long-running agents can spend heavily on repeatedly processing information they already encountered.