ThoughtDAG is an open-source, local-first workspace that treats conversation edges as the mechanism for controlling model context. Each question-and-answer exchange functions as a node, and only upstream wired nodes are serialized into the next model request.

  • Removing an edge eliminates that branch from the actual input rather than just the visualization.
  • The author is running a pilot benchmark with three perturbation families: irrelevant plausible branches, outdated assumptions, and conflicting evidence.
  • Comparisons use fixed settings across clean evidence (C0), polluted context (C1), explicit graph removal (C2), and automatic summarization (C3).
  • The tool supports Ollama for local inference and arbitrary OpenAI-compatible endpoints, storing all data locally.

The author is seeking feedback from users running local models regarding suitable model/context combinations, missing real-world contamination cases, and the utility of explicit pruning versus retrieval in smaller context windows.