Multi-agent LLM systems typically communicate in text, but that is lossy and wasteful: the sender throws away the reading it already did. A growing literature instead transmits the KV cache, the natural carrier of that reading, so the receiver can start from a richer state.

The new arXiv paper, "Close Enough is Nowhere Near: Fidelity-Consumability Decoupling in KV Caches," studies this approach and argues for separating two properties of a transmitted cache. Fidelity measures how closely the cache matches the original context; consumability measures how well the receiving model can actually use it. The title suggests that approximate fidelity is not enough on its own.

The abstract is brief, so the full findings are not yet available from the announcement. But the framing points to a design tension: a cache that is highly faithful to the sender's context may not be easily consumable by the receiver, and optimizing one property may not optimize the other.