CacheBack lets a receiving agent describe what information it needs, then uses the sender’s attention weights to select and transmit only relevant KV-cache positions. It requires no training and avoids both text generation and full-cache transfer.
On FanOutQA with Qwen 3, it removed 75% of the incoming state, improved accuracy by 14.7 percentage points, and reduced median task-completion latency by 3.2x versus text communication. Similar gains appeared across dense Transformers, Mamba-attention hybrids, sliding-window attention, and both parallel and sequential agent topologies.