Agent Memory: First Systems Characterization of Stateful Long-Horizon LLM-Agent Workloads

arXiv·high signal

Researchers present the first systems-level characterization of agent memory architectures spanning flat retrieval, LLM-mediated extraction, consolidating fact stores, and agentic control flows. The paper measures the runtime and storage behavior of agents that persistently store, retrieve, and update memory across sessions — directly relevant to anyone building long-running agent pipelines. It surfaces the system implications (latency, cost, retrieval overhead) that memory frameworks impose at scale.