Valkey 9.2 launches alongside a shared cache for AI inference

The Valkey project released version 9.2 of its open-source datastore on November 19 alongside a distributed cache designed to reduce repeated work during AI inference.
The project’s announcement describes a cache shared across a GPU fleet. Inference engines can reuse intermediate computations from recurring prompt content, including tool definitions and conversation history, instead of processing those sections from scratch. Teams serving open-weight models on their own infrastructure can connect through LMCache, which integrates with vLLM and SGLang.
Valkey 9.2 adds Path Hash, a data type that can find the longest matching prefix and locate cached results. Valkey can serve as the index while larger data remains in GPU or host memory. The release notes also list opt-in background snapshots that do not require forking the datastore process.
The shared cache addresses a recurring cost in agent workflows: the same context often reaches different workers. Actual savings depend on how often those workers can reuse cached computations.
Sources and citations
Revision history
Published .