Reproduce “AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM 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
Xiangchen Cheng, Yunwei Jiang, Jianwen Sun, Zizhen Li, Chuanhao Li, Xiangcheng Cao, Yihao Liu, Fanrui Zhang and 2 more
A new framework for long-horizon LLM agents introduces a bounded-memory contract, preventing unbounded context growth by using per-decision typed retrieval. This approach demonstrates enhanced performance and operational efficiency for agents navigating the complex, stochastic Slay the Spire 2 game environment.
Key insights
- Explicitly structured and typed memory layers, particularly L5 strategic skills, provide substantial performance improvements for LLM agents in long-horizon, stochastic environments.
- A bounded-memory contract significantly reduces token costs and execution time while improving agent performance compared to traditional accumulating-context approaches.
- The effectiveness of a pre-trained memory stack (L4+L5) can vary considerably across different LLM backbones, highlighting the importance of backbone compatibility in agent design.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.