Subconscious Cache: Reliably Capture Your Agent Context

A new technique called Subconscious Cache improves the efficiency of AI agents by addressing the limitations of standard prefix caching in LLM inference systems. When an agent's context is compacted, traditional systems must re-encode large token spans, but Subconscious Cache reuses both prefix and suffix states. This preserves information from pruned sections, keeping it in the model's "subconscious" reasoning. The result is a significant reduction in latency, improved system throughput, and better performance on long-horizon reasoning tasks.