ThoughtDAG是一款针对大语言模型对话的上下文管理工具，通过可视化和可编辑的图形界面让用户精细控制对话历史中哪些信息被纳入下一个请求[1]。该工具的核心机制在于使用图形结构表示对话上下文，用户可以直接编辑和删除特定的上下文边以调整模型的输入范围。
工具演示表明，删除单条无关的上下文边（例如与对话核心无关的晚餐建议）可以减少对话上下文的规模，删除橙色边后的案例中减少了47个tokens的上下文[1]。更关键的是，仅通过修改一条上下文边，在保持提示词相同的情况下，模型的答案会发生更新[1]。ThoughtDAG还提供了三层可重现性机制来确保结果的可靠性，包括记录数据库版本、使用独立审阅者和引入第三方冲突解决[1]。该工具具备可见（Visible）、可编辑（Editable）和可检查（Inspectable）三个核心特性[1]。
ThoughtDAG is a tool designed to give users fine-grained control over conversation context in large language model interactions through a visual, editable graph interface [1]. The platform allows users to manage which information from conversation history is included in subsequent requests, enabling both precise context management and pollution detection [1].
The tool operates by representing conversation elements as nodes and edges within a graph that users can modify directly [1]. By removing specific context connections—such as irrelevant information like dinner suggestions—users can see immediate changes in model responses [1]. In one demonstrated example, deleting a single orange edge reduced the context by 47 tokens, and changing only that one connection caused the model's answer to update with the same prompt [1].
ThoughtDAG emphasizes three core design principles: visibility (Visible), editability (Editable), and inspectability (Inspectable) [1]. The platform incorporates three layers of reproducibility mechanisms, including version tracking in a recorded database, the use of independent reviewers, and third-party conflict resolution [1].