AI agents win at Slay the Spire 2 after researchers replace growing chat logs with structured memory

2026-07-12 · Source: The Decoder · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, medium

Summary

The AgenticSTS project, developed by Alaya Lab and Shanghai Jiao Tong University, introduces a novel memory architecture for AI agents, enabling them to succeed in the digital card game Slay the Spire 2 where frontier models previously failed. Instead of appending to an ever-growing chat log, AgenticSTS rebuilds each decision from five organized memory layers: fixed protocol instructions (L1), state schemas (L2), retrieved game rules (L3), summaries of previous runs (L4), and strategy skills (L5). This approach significantly improved performance; activating the L5 skill library doubled the win rate from 3 out of 10 to 6 out of 10 games at A0 difficulty. Furthermore, AgenticSTS demonstrated superior efficiency against public agents STS2MCP and CharTyr, winning 6 of 10 games compared to their 0 of 5, while using 66 to 90 times fewer tokens per point. Its prompt size remained constant at approximately 5,000 tokens, contrasting with competitors' prompts that grew to 527,000 tokens.

Key takeaway

For Machine Learning Engineers designing AI agents for complex, multi-step environments, you should prioritize structured, fixed-size memory architectures over traditional growing chat logs. This approach dramatically improves agent performance and reduces token costs by preventing "context rot." Implement distinct memory layers for protocols, state, rules, run summaries, and skills to enhance decision-making and efficiency. Your agent's accumulated knowledge might be model-specific, so test transferability carefully.

Key insights

Structured, fixed-size memory layers significantly enhance AI agent performance and efficiency over growing conversational logs.

Principles

• Fixed-size, layered memory prevents context rot and improves agent performance.
• Separating memory layers allows pinpointing specific drivers of agent behavior.
• Knowledge accumulated by one model may not transfer effectively to others.

Method

Rebuild each agent decision prompt from five distinct memory layers: fixed protocol instructions (L1), state schemas (L2), retrieved game rules (L3), summaries of previous runs (L4), and situation-specific strategy skills (L5).

In practice

• Implement a fixed-size, layered memory architecture for complex agent tasks.
• Store tactical rules in a dedicated skill library for recurring situations.
• Publish datasets of agent runs and memory snapshots for comparative research.

Topics

• AI Agents
• Structured Memory
• Context Management
• Slay the Spire 2
• Token Efficiency
• Large Language Models

Code references

• Gennadiyev/STS2MCP
• CharTyr/STS2-Agent

Best for: Research Scientist, AI Engineer, NLP Engineer, AI Scientist, Machine Learning Engineer, AI Architect

Related on AIssential

• Agent memory resolved? — A git-like memory framework significantly enhances coding agent performance and context persistence across sessions.
• Your AI Doesn’t Forget. It Just Runs Out of Space. — AI models lack inherent memory; applications must actively manage context windows to simulate long-term conversation recall.
• AI Agents Have Amnesia. A Bigger Context Window Won’t Cure It. — AI agents require a dedicated, architected memory layer to overcome inherent amnesia, as context windows are buffers, not durable truth models.
• Context Engineering for AI Agents: Part 2 — Effective AI agent performance hinges on precise context management, minimizing irrelevant information and optimizing tool access.
• AI Agents Have Four Kinds of Memory, Not One — AI agents need four specialized memory types and two active management habits for true long-term recall, beyond just larger capacity.
• Graph-Validated Memory Architecture: Enhancing Contextual Accuracy and Retrieval Speed in Agentic… — The Graph-Validated Memory Architecture resolves the latency-accuracy trade-off in agentic systems by combining fast semantic caching with knowledge graph validation.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Decoder.