ValkeyConf 2026

Agentic AI applications are GPU intensive and expensive to run. KV caching allows reuse of computed work for repeated context to reduce latency and costs by up to 80%. This talk explains how KV caching works and walks through a production architecture using vLLM as the inference engine and Valkey as a distributed cache. The talk examines performance implications of cache tuning with LMCache and recent innovations in Valkey that break the ceiling on this performance. You will leave with practical implementation guidance and a framework for evaluating when a database-backed KV cache tier improves your AI application.