Persistent KV Cache Across Sessions: True Stateful Memory for Large Language Models

We demonstrate persistent KV cache across sessions in large language models: the internal attention state of a transformer is serialized to disk, reloaded in a new process, and used directly for inference without reprocessing the original context. Using Mistral 7B (bfloat16) on a Tesla T4 GPU, an 84-token conversational context encoded into a 10.52 MB file enables accurate recall of user identity and research topics across session boundaries — with no text history passed to the model. The model correctly identifies the user by name using only the deserialized KV cache as input. This establishes a proof-of-concept for stateful AI memory: persistent, model-native, and free of retrieval overhead. Seventh paper in a series on KV cache optimization for consumer GPU inference.