AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents
The authors build a Slay the Spire 2 benchmark where an agent must make hundreds of connected decisions with tight memory limits. They swap the usual endlessly growing context for a "bounded contract" that pulls only typed, relevant records into each step. Use this testbed to stress‑test your agent memory designs instead of guessing from toy games.
Xiangchen Cheng, Yunwei Jiang