Reproduce “Persistent Memory in Multi-Agent LLM Inference: What It Costs, What It Buys, and When You Can Tell” 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
Hochan Son, Kyungdoe Han, Jaehan Koh, Xiaowu Dai, Wenlu Xu, Guang Cheng
The chain-of-agents system tests whether distributing evidence across calls reduces peak KV cache compared with single-pass and retrieval-augmented systems. It passes distilled information between agents, so no one call holds all evidence. In pooled HotpotQA and LB-HotpotQA comparisons, it measured 14.3 MiB peak cache, versus 35.5 and 35.3 MiB. This reduces peak cache, not total KV traffic or device occupancy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.