A long search produces a lot of text, but text is not the same as memory. The agent may have already found a decisive fact twenty actions ago. Once that fact falls outside the active window, it can repeat the search, forget a constraint, or accept a candidate it had previously ruled out.长程搜索会产生大量文本,但文本不等于记忆。智能体可能在二十个动作前就找到了关键事实;当这段信息离开当前窗口,它又会重复检索、遗忘约束,甚至重新接受早已排除的候选。

The solution is not simply to keep every token in view. The agent needs a smaller object that says what the task currently knows. In ClueFlow, that object is a persistent state with evidence records and candidate records.解决办法并非简单地让所有 token 永远留在视野里。智能体需要一个更小的对象,说明任务目前已经知道什么。在 ClueFlow 中,这个对象由持久的证据记录与候选记录构成。

State is an accountable claim状态是一种需要负责的判断

An evidence record should say more than “I saw this.” It should identify the source and capture a fact in a form that can be added, revised, or discarded. A candidate record has a different job: it connects a possible answer to supporting evidence, missing requirements, and contradictions.一条证据记录不能只说“我看到了这个”。它需要标明来源,以可新增、修订或废弃的形式保存事实。候选记录承担另一项职责:把可能的答案与支持证据、尚缺的条件和已有矛盾连接起来。

The separation matters. Evidence is about what a source supports; a candidate is about what the current evidence allows the agent to conclude. Mixing the two makes it harder to see whether a wrong final answer came from bad retrieval, a bad state update, or a bad final decision.两者分离很重要。证据关乎来源支持什么,候选关乎当前证据允许智能体得出什么结论。混在一起,就难以判断错误答案究竟源自检索、状态更新,还是最终决策。

From a browser action to a learning signal从浏览动作到训练信号

A provenance chain connects the source action, the state update, and the final conclusion. This gives the agent an operational discipline during inference: when a candidate is nearly ready, it can see exactly which requirement remains unresolved. It also gives training a finer unit of analysis than the terminal reward.来源链把浏览动作、状态更新和最终结论连起来。这让智能体在推理时更有章法:候选即将完成时,能看清还有哪项条件未核实。它也让训练拥有比终局奖励更细的分析单位。

A run can fail overall and still contain a good source discovery or a correct state update. Preserving those steps helps training reward the parts of a trajectory worth repeating, while keeping the final outcome as a separate signal. That distinction is central to ClueFlow's hybrid supervision and turn-level credit design.一次运行整体失败,也可能包含正确的来源发现或状态更新。保留这些局部正确动作,有助于训练重复值得重复的部分,同时让最终结果保持为独立信号。这是 ClueFlow 混合监督与逐轮信用设计的核心区分。

The broader lesson is that agent memory is part of the policy. It changes what the model can see, what it can verify, and which decisions can later be explained.更一般的启示是:智能体记忆本身就是策略的一部分。它决定模型能看见什么、能核实什么,以及哪些决策最终能够被解释。