Hacker News
Show HN: ThoughtDAG Makes an LLM Conversation's Context an Editable Graph You Can Prune Before Sending
ThoughtDAG, on HN at 49 points, argues that chat UIs show what was said but not which history actually enters the next request, and replaces the linear transcript with a directed acyclic graph where nodes are Q&A exchanges and edges are context connections. Users branch to explore alternative interpretations without overwriting, prune specific edges to exclude them from the request while keeping them visible, merge reasoning paths, and preview the exact payload before generation — then delete an edge and regenerate to confirm the context truly changed. MIT-licensed with source on GitHub, a Cloudflare Workers web demo, and a v0.3.13 desktop build for macOS, Windows, and Linux. No benchmarks or paper are offered.
↳ Follow the thread