精选理由
LangChain 推荐了 Candidly 的 IO-HMM 框架，把对话拆成用户和智能体两半，状态管理更干净。
LangChain 在推文中介绍了 @GetCandidly 提出的 IO-HMM（输入-输出隐马尔可夫模型）框架。该框架将对话的每一轮明确拆分为用户行为（可观测信号）和智能体行为（可控输入）。用户行为用于读取状态，智能体行为则转移状态。这种设计让对话状态管理更清晰。
原文 · LangChain
Every turn of a conversation has two pieces. 👤 What the user did: observable signals that tell you ...
Every turn of a conversation has two pieces. 👤 What the user did: observable signals that tell you where the conversation is 🤖 What the agent did: controllable inputs that condition where it goes next @GetCandidly 's IO-HMM keeps these separate by design. 👤 User behavior reads the state 🤖 Agent behavior shifts it 💬 2 🔄 0 ❤️ 5 👀 1071 📊 3 ⚡