Reproduce “SkillMem: Compact External Skill Memory for Latent Multi-Agent Reasoning” 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
Yuxuan Zhu, Mohammad Mohammadi Amiri, Pin-Yu Chen, Keerthiram Murugesan
SkillMem stores procedures as compact external memories for multi-agent systems, rather than repeatedly supplying their text. A shared reader uses each agent’s role and hidden state to access memory; the backbone stays frozen while the memory and reader learn the procedure’s behavioral effects. On GPQA-Diamond, Qwen3-4B reaches 35.63% accuracy versus 17.97% for LatentMAS, but gains are not consistent across model sizes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.