Reproduce “Hippo: Training-Free Agent Self-Evolution via Hippocampal Replay” 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
Yutong Song, Jiang Wu, Yuxuan Ren, Tony Qin, Baoyuan Wang
Hippo helps agents revise external memory by comparing multiple attempts at a task, diagnosing errors and misleading successes, and updating guidance with conditions for when it applies. With GPT-5.6 Sol, it raises WebArena overall task success to 72.9%, from 64.2% without memory. This uses eight attempts per task and does not show lower cost or indefinite gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.