Reproduce “MemoryDojo: Benchmarking Poisoning Through LLM Agent Memory” 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

David H. Yang, Blake Bullwinkel, Amanda Minnich, Mohammad Mohammadi Amiri, Giorgio Severi

MemoryDojo evaluates whether poisoned tool observations can affect later tasks through agents’ memory systems. It follows memory writes, retrievals, and subsequent behavior across five complete integrations. In same-domain GPT-5.6-Luna experiments, A-MEM BRT rises from 21.02% to 44.01% under payload optimization, while RSR remains similar. This suggests retrieval alone does not determine behavioral reach in these sandboxed settings.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.