Infrastructure & deployment
ReCache: Efficient KV Cache Reuse and Compression for Tool-Augmented LLM Agents
ReCache tackles why standard prefix caching misses repeated tool and skill schemas when agents combine them in different orders.
Why it matters: For agents that repeatedly use tool or skill schemas in changing combinations, ReCache offers a way to reuse schema representations while reducing inference costs. The reported gains concern specific invocation, KV-memory, and attention measures—not overall task accuracy or total deployment memory.