🛰️ Daily AI Frontier
‹ back to 2026-08-23

ReCache: Efficient KV Cache Reuse and Compression for Tool-Augmented LLM Agents

Research Efficiency & Systems

Ranking

Overall 87
Content 95
Popularity 69

Observed public metrics from 1 member.

Merged summary

TL;DR - ReCache makes tool-augmented LLM agents more efficient by caching tool and skill schema representations independently, allowing reuse even when resources appear in different combinations or orders. It substantially reduces latency and KV-cache memory with minimal invocation-quality loss.

  • Resource-wise attention creates composition-invariant KV blocks by removing cross-resource interactions and using resource-local positions.
  • Selective layer/head-group routing and structural-semantic pruning retain only invocation-critical schema information.
  • Resource-wise attention achieved 82.3% Inv-F1 versus 82.4% for dense invocation, with a 3.655× time-to-first-token speedup.
  • The complete framework reduced allocated KV-tensor memory by 92.43% and accelerated attention by 1.423×.

Sources (1)

ReCache: Efficient KV Cache Reuse and Compression for Tool-Augmented LLM Agents

arXiv cs.CL Yichu Fang, Sitong Wei, Haozhe Hu, Xiaoyu Shen 2026-08-20 arXiv:2608.19662
Public signals Semantic Scholar citations 1 · Semantic Scholar influential citations 0
Providers: Hugging Face · N/A OpenAlex · N/A Publisher · N/A Semantic Scholar · Citations 1 · Influential citations 0 X · N/A Fetched 2026-09-14 14:18:45.516680 UTC

TL;DR - ReCache makes tool-augmented LLM agents more efficient by caching tool and skill schema representations independently, allowing reuse even when resources appear in different combinations or orders. It substantially reduces latency and KV-cache memory with minimal invocation-quality loss.

  • Resource-wise attention creates composition-invariant KV blocks by removing cross-resource interactions and using resource-local positions.
  • Selective layer/head-group routing and structural-semantic pruning retain only invocation-critical schema information.
  • Resource-wise attention achieved 82.3% Inv-F1 versus 82.4% for dense invocation, with a 3.655× time-to-first-token speedup.
  • The complete framework reduced allocated KV-tensor memory by 92.43% and accelerated attention by 1.423×.
item →