ThoughtDAG creator shares exercise: disconnect a discussion's context path, regenerate, and compare the answer

Products·building

ThoughtDAG: open-source conversation canvas for branching AI chats with selective context

ThoughtDAG is an open-source conversation canvas where you can branch AI discussions and wire chosen prior paths as context into the next answer. A history index finds candidate excerpts from past conversations, an optional step called Jev ranks relevance, and your regular LLM develops the answer. It targets people who work through research questions across sessions and want decisions carried forward without dragging entire old conversations along.

ThoughtDAG is an open-source conversation canvas where you can branch AI discussions and wire chosen prior paths as context into the next answer. A history index finds candidate excerpts from past conversations, an optional step called Jev ranks relevance, and your regular LLM develops the answer. It targets people who work through research questions across sessions and want decisions carried forward without dragging entire old conversations along.

Products·building

ThoughtDAG: User-editable DAG for LLM context instead of automatic memory

ThoughtDAG is an experimental system that replaces automatic LLM memory retrieval with a visible, user-editable directed acyclic graph (DAG) of conversation nodes. Each question/answer exchange is a node, and edges determine which nodes are included in the next model request. Users can delete edges to remove branches from actual context, giving them direct control over what the LLM sees.

ThoughtDAG is an experimental system that replaces automatic LLM memory retrieval with a visible, user-editable directed acyclic graph (DAG) of conversation nodes. Each question/answer exchange is a node, and edges determine which nodes are included in the next model request. Users can delete edges to remove branches from actual context, giving them direct control over what the LLM sees.

Products·launched

ThoughtDAG: graph-based LLM canvas where edges define context

ThoughtDAG is an open-source Ollama canvas that uses a directed acyclic graph (DAG) structure where each node is a question-answer exchange and edges determine the context sent to the LLM. When a user asks from a node, only its upstream ancestors are included in the prompt, making context explicit and editable. It solves the problem of linear chat interfaces that hide context, giving users fine-grained control over what the model sees.

ThoughtDAG is an open-source Ollama canvas that uses a directed acyclic graph (DAG) structure where each node is a question-answer exchange and edges determine the context sent to the LLM. When a user asks from a node, only its upstream ancestors are included in the prompt, making context explicit and editable. It solves the problem of linear chat interfaces that hide context, giving users fine-grained control over what the model sees.

Products·launched

ThoughtDAG: branch side questions from chat answers without losing context

ThoughtDAG is an open-source chat tool that lets users select a phrase in an LLM answer and explore it on a branch, then connect parts into the next question. It supports Ollama and OpenAI-compatible APIs, demonstrated with DeepSeek. It solves the problem of exploring tangents without disrupting the original conversation, for developers and researchers who chat with LLMs.

ThoughtDAG is an open-source chat tool that lets users select a phrase in an LLM answer and explore it on a branch, then connect parts into the next question. It supports Ollama and OpenAI-compatible APIs, demonstrated with DeepSeek. It solves the problem of exploring tangents without disrupting the original conversation, for developers and researchers who chat with LLMs.

Developments·Paper

Pilot benchmark shows context pollution survives source deletion in multi-turn LLM conversations

A researcher introduced a pilot benchmark for context pollution in multi-turn LLM conversations, showing that deleting or editing the original erroneous message may not repair the conversation because later turns can repeat or derive from the polluted information. The work frames this as a context-engineering problem rather than a hidden-reasoning problem.

A researcher introduced a pilot benchmark for context pollution in multi-turn LLM conversations, showing that deleting or editing the original erroneous message may not repair the conversation because later turns can repeat or derive from the polluted information. The work frames this as a context-engineering problem rather than a hidden-reasoning problem.