Attention-Derived Dependency Cut for Independent KV-Cache Reuse in Agent Memory

Long-term agent memory grows continuously with interaction, making full-memory prefill increasingly costly in terms of time to first token (TTFT) and redundant computation. KV-cache reuse could avoid repeatedly prefilling the accumulated memory, but KV caching relies on exact prefix matching. When memory is updated, reorganized, or assembled from reusable units, prefix mismatches prevent direct reuse of existing KV caches and force the model to recompute a long memory prefix. Existing methods largely assume that memory blocks are already given and address the prefix-matching limitation by concatenating their KV caches and then repairing the resulting mismatch. We therefore argue that the primary question for agent-memory KV reuse is where to segment the memory. We propose AD-CUT, an Attention-derived dependency Cut method for training-free, independence-oriented segmentation. AD-Cut computes attention-derived dependency signals over sentence-level atomic units, including an incoming signal Ain and an outgoing signal Aout, and selects boundaries where both signals are low. The signals exhibit stable positive associations with intervention-derived KV and NLL sensitivity, with Spearman ρ ranging from 0.615 to 0.868 and context-cluster bootstrap intervals remaining positive, supporting their use as predictive proxies. In block-KV serving for agent QA, AD-Cut achieves a 16.17× Online TTFT speedup with a QA Quality Retention of 0.7675. It also improves task quality over naive sentence-level reuse in a controlled agent-memory evaluation. These results show that AD-Cut substantially reduces online computation while mitigating the quality loss caused by KV-cache reuse.