在刪除聊天記錄後，Slay the Spire 2 智能體勝率從0飆升至60%
研究人員用結構化內存取代了不斷增長的聊天日誌，使智能體在 Slay the Spire 2 中勝率達到 60%，證明上下文管理比模型尺寸更重要。
frontier models failed 100% of attempts at slay the spire 2, but the agenticsts architecture achieved a 60% win rate by discarding chat logs in favor of structured memory slots. the system uses five layers—protocol, state, rules, run summaries, and skills—to rebuild the prompt for each decision without context bloat. this approach proved that episodic learning and skill libraries are more critical than raw model power for complex planning tasks.