Reproduce “Delivered, Not Used: Evaluating Agent Memory at Its Point of Use” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Changran Xu, Zeju Li, Qiang Xu
MemUse evaluates delivered memory use in recorded agent trajectories. It compares each delivery with the agent’s prior context, then traces subsequent actions and consequences. In SWE-bench Verified runs, 89 of 383 episodes with beneficial memory use failed, while 5 of 16 with harmful use passed. These judge-assessed observations do not establish memory’s causal effect on task success.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.