MemAware – AI 에이전트가 "내가 뭘 알고 있는지"를 아는지 측정하는 벤치마크

MemAware exposes a critical flaw in existing AI agent memory benchmarks—they only measure retrieval capacity, not genuine memory understanding. By quantifying how vector and keyword search fail on cross-domain context recall (0.7% accuracy on hard tasks), this work reveals structural limitations in current RAG-based memory systems like ChatGPT Memory and MemGPT, challenging the assumption that better search solves agent memory problems.