Workshops Archive

The Agentic Memory Hierarchy: Cross-State I/O Characterization for a Unified Substrate


Workshop: AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC

Authors: Avinash Kumar Maurya (Argonne National Laboratory (ANL)), Bogdan Nicolae (Argonne National Laboratory), and M. Mustafa Rafique (Rochester Institute of Technology)

Abstract: LLM agents now run long, multi-agent, multi-user workloads whose state does not fit a single store. An agent produces four kinds of state on one memory-storage hierarchy: the model context (KV cache), execution state of tool calls and checkpoints, cognitive memory of facts and skills, and bulk artifacts and embeddings. Each sits behind a separate, uncoordinated substrate, so lifetimes are uncorrelated, repeated content is unrecognized, evictions collide over one memory budget, and provenance is lost. We present the first cross-state I/O characterization, via an extract-normalize-replay method over real agent, serving, memory, and HPC I/O traces and the static footprint of four scientific agents. The states occupy distinct size regimes, span two stress axes, and contend in most windows; over half of KV prefix blocks are reusable, yet recency eviction captures only a third. Each finding maps to a unified-substrate requirement; we release the schema, extractors, and replay benchmark.


Back to AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC Archive Listing

Back to Full Workshop Archive Listing