Hi everyone, I am the author of ThoughtDAG, an open-source, local-first workspace built around one testable rule: wires are the context.
Each question-and-answer exchange is a node. When a user asks from a node, only its wired upstream nodes are serialized into the next model request. If an edge is removed and the same prompt is regenerated, that branch disappears from the actual input rather than only from the visualization.
I built it for research workflows where one question branches into several hypotheses, source documents, extracted passages, and counterexamples. The user can keep exploratory branches visible while deciding which evidence should continue into the next inference.
I am currently running a small context-control benchmark pilot with three perturbation families:
The required comparison uses the same fixed model and settings:
I would especially value feedback from people running local models:
ThoughtDAG supports Ollama for fully local inference and arbitrary OpenAI-compatible endpoints. The desktop app stores canvases, documents, and keys locally.
Project source and setup: GitHub - chenxiachan/thoughtdag: Your thinking deserves a map: an infinite canvas where LLM conversations grow into an editable thought graph. Wires are the context. · GitHub
This is an early prototype and the benchmark is still running. I am more interested in failure cases and critiques of the interaction model than in general project promotion.