Today’s top-tier models often fail at complex, multi-step tasks—not because they lack "intelligence," but because of the flawed way they handle dialogue history. A recent study by Alaya Lab and Shanghai Jiao Tong University (SJTU), using the game Slay the Spire 2 as a benchmark, illustrates the problem: traditional chat logs that bloat with every move inevitably dilute the model's attention. In the AgenticSTS project, researchers documented a fatal flaw—agents powered by current LLMs failed every single run when they simply "appended" past observations and tool calls into the context window.
The solution proposed by the AgenticSTS team is a radical shift in strategy. Instead of a linear dialogue history, they implemented a fixed catalog of five structured slots—ranging from protocol instructions to a "skill library" containing tactical rules. This architecture keeps the prompt concise regardless of how long the session lasts.
Shifting from "episodic accumulation" to "architectural memory" allowed the agent to reach difficulty levels A6–A8, while models without active memory stalled at the start.
Activating the skill library doubled the win rate—climbing from 3 to 6 victories out of 10. The prompt remains compact throughout the session, regardless of duration. Decision-making accuracy improves by filtering out noise within the context.
For executives leading AI transformations, there is a pragmatic economic takeaway here. As businesses move toward autonomous agents in ERP systems or supply chain management, trying to "chat with the past" is a dead end. The only way to ensure reliability and cost control is to transition to structured state queries.
If your AI agents still rely on endless logs to track workflows, you are overpaying for a system that degrades as it works. Success in complex business logic requires moving away from the obsession with context window size and toward a fixed-slot memory architecture. It is the only way to maintain precision without wasting tokens or letting your AI turn into a forgetful bureaucrat.