AI News AI资讯 6h ago Updated 2h ago 更新于 2小时前 48

AI agents win at Slay the Spire 2 after researchers replace growing chat logs with structured memory 研究人员用结构化记忆替换不断增长的聊天日志,AI代理在《杀戮尖塔2》中获胜

AgenticSTS replaces growing context logs with a structured five-layer memory architecture to prevent "context rot" and maintain performance in long-horizon tasks. The system achieves a 60% win rate in Slay the Spire 2 by utilizing an L5 skill library, doubling the baseline performance of frontier models without memory. Compared to traditional agents, AgenticSTS reduces token usage by 66-90x and latency by 4x, keeping prompt size constant regardless of game duration. Memory transferability is lim AgenticSTS项目提出了一种五层固定槽位的记忆架构,取代了传统LLM代理中不断增长的对话日志,解决了“上下文腐烂”问题。 在《杀戮尖塔2》测试中,启用L5技能库使AI代理的胜率从30%提升至60%,证明了结构化记忆对长期决策的重要性。 相比采用增长日志模式的公开代理,AgenticSTS将每回合Token消耗降低约66-90倍,显著减少了延迟和计算成本。 跨模型迁移实验显示,记忆内容具有强模型依赖性,一个模型积累的记忆难以直接提升其他模型的 performance。

68
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • AgenticSTS replaces growing context logs with a structured five-layer memory architecture to prevent "context rot" and maintain performance in long-horizon tasks.
  • The system achieves a 60% win rate in Slay the Spire 2 by utilizing an L5 skill library, doubling the baseline performance of frontier models without memory.
  • Compared to traditional agents, AgenticSTS reduces token usage by 66-90x and latency by 4x, keeping prompt size constant regardless of game duration.
  • Memory transferability is limited; strategies learned by one model do not effectively transfer to others, indicating strong model-specific dependencies in learned behaviors.

Why It Matters

This research addresses the critical scalability bottleneck of "context rot" in agentic AI, demonstrating that efficient memory management is more impactful than raw model capability for complex, multi-step reasoning tasks. By proving that structured, bounded memory can significantly outperform unbounded context windows, it provides a practical blueprint for reducing inference costs and improving reliability in real-world applications requiring long-term state retention.

Technical Details

  • Five-Layer Memory Architecture: The agent constructs prompts from five distinct slots: L1 (fixed protocols), L2 (current state schemas), L3 (retrieved rules), L4 (episodic summaries of past runs), and L5 (strategic skills).
  • Skill Library Implementation: L5 stores tactical rules for recurring situations; enabling this layer increased win rates from 30% to 60% on the lowest difficulty setting.
  • Performance Benchmarks: Tested on Slay the Spire 2, where frontier models previously failed completely. The agent reached higher difficulties (A6-A8) only when inter-run memory (L4/L5) was active.
  • Efficiency Metrics: Maintained a consistent ~5,000 token count per turn, whereas competing agents with growing logs reached up to 527,000 tokens, causing significant latency increases.

Industry Insight

  • Shift from Context to Structure: Developers should prioritize designing explicit memory structures and retrieval mechanisms over relying on expanding context windows to manage long-running agent sessions.
  • Cost Optimization: Implementing bounded memory architectures can drastically reduce API costs and latency, making complex agentic workflows economically viable at scale.
  • Model-Specific Memory: Since learned memory does not easily transfer between different model architectures, teams should expect to fine-tune or retrain memory components for each specific model deployment rather than treating memory as a portable asset.

TL;DR

  • AgenticSTS项目提出了一种五层固定槽位的记忆架构,取代了传统LLM代理中不断增长的对话日志,解决了“上下文腐烂”问题。
  • 在《杀戮尖塔2》测试中,启用L5技能库使AI代理的胜率从30%提升至60%,证明了结构化记忆对长期决策的重要性。
  • 相比采用增长日志模式的公开代理,AgenticSTS将每回合Token消耗降低约66-90倍,显著减少了延迟和计算成本。
  • 跨模型迁移实验显示,记忆内容具有强模型依赖性,一个模型积累的记忆难以直接提升其他模型的 performance。

为什么值得看

该研究揭示了当前AI智能体在长程任务中面临的核心瓶颈并非模型能力本身,而是记忆管理效率低下导致的性能衰减和高昂成本。它为构建高效、可扩展的长期自主智能体提供了新的架构范式,即通过结构化、分层的外部记忆替代无限增长的内部上下文窗口。

技术解析

  • 五层记忆架构:AgenticSTS将记忆分为五个独立槽位:L1为固定协议指令,L2为当前有效动作的状态模式,L3为检索到的游戏规则,L4为跨局次的运行摘要( episodic memory),L5为特定情境下的策略技能库。每次决策时,提示词由这些固定槽位重新构建,而非追加历史日志。
  • 性能与成本对比:在A0难度下,无记忆层胜率为30%,仅开启L5技能库后胜率翻倍至60%。与STS2MCP和CharTyr等公开代理相比,后者在长游戏中单轮调用可达527,000 Tokens,而AgenticSTS稳定在约5,000 Tokens,且耗时仅为前者的四分之一。
  • 记忆迁移局限性:实验测试了将Gemini 3.1 Pro积累的记忆栈直接传递给Qwen 3.6-27B和Deepseek V4-Pro,结果显示Qwen得分提升84.5%,但Deepseek反而下降18.1%,且均未获胜,表明记忆内容与创建它的模型紧密耦合,通用性较差。
  • 基准测试环境:以《杀戮尖塔2》为测试床,因其规则完全文本化、随机性高且单次游戏包含数百次决策,能清晰暴露不同记忆架构在处理长程依赖时的差异。

行业启示

  • 从“增长上下文”转向“结构化存储”:随着智能体执行任务长度增加,传统的ReAct或Reflexion等追加日志模式将面临严重的效率瓶颈。行业应优先研发类似AgenticSTS的分层、可检索的外部记忆机制,以维持推理速度和准确性。
  • 记忆模块的标准化与隔离:将协议、状态、规则、经验和技能分离存储,不仅提升了效率,还使得研究人员能够精确归因于哪一层记忆带来了性能提升,这为优化智能体行为提供了可解释的工程路径。
  • 关注记忆的可移植性与泛化能力:当前记忆技术存在强烈的模型绑定效应,限制了知识的复用。未来研究需解决如何构建模型无关的通用记忆表示,或开发更高效的记忆迁移和适配算法,以降低多模型协作的成本。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Agent Agent Research 科学研究 Gaming 游戏 LLM 大模型