Reproduce “HaluMem: Evaluating Hallucinations in Memory Systems of Agents” 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

Ding Chen, Simin Niu, Kehang Li, Peng Liu, Xiangping Zheng, Bo Tang, Xinchi Li, Feiyu Xiong and 1 more

HaluMem evaluates long-term conversational memory by scoring fact extraction, memory updating, and question answering separately. It uses simulated user histories and annotated memory points to identify where errors occur, rather than relying only on final answers. On HaluMem-Medium, MemOS reaches 74.07% extraction recall and 67.23% QA accuracy, versus Mem0’s 42.91% and 53.02%; these system comparisons do not establish causation.

Key insights

  • Hallucinations are not isolated events but can accumulate during early memory operations (extraction, updating) and propagate, amplifying their impact on downstream tasks like question answering.
  • Existing memory systems struggle to balance comprehensive memory coverage with high extraction accuracy, frequently omitting crucial information while also extracting irrelevant or fabricated content.
  • Dialogue addition, encompassing memory extraction and updating, represents the primary computational bottleneck for current memory systems, requiring significantly more time than memory retrieval.

Summary by alphaXiv.

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