ThoughtDAG: An Editable Context Graph for LLM Conversations

ThoughtDAG is a desktop app that replaces the linear chat log with an editable context graph. Instead of hiding which parts of a conversation feed the next request, ThoughtDAG shows the structure — and lets you change it. The core idea: the graph is the context.

Wires as context

In ThoughtDAG, nodes are sources and messages; edges (wires) determine what gets included in the next LLM request. You can clip a passage from a PDF, ask a question directly from a source node, or spawn an unrelated branch for a detour — all while keeping the main path clean.

• Branch: explore an alternative interpretation without overwriting your current path.
• Prune: keep a useful detour on the canvas but exclude it from the next request.
• Merge: bring selected evidence and reasoning paths back together in one answer.
• Inspect: preview exactly what the model will receive before generation — including order and token count.

One rule governs everything: Wires are context. No hidden memory selector. Deleting an edge actually removes that branch from the prompt — it's not just a visual aid.

Visible diffs and reproducible answers

The interface shows a token diff when you change edges. For example, removing an unrelated detour about dinner saves 47 tokens. And because the prompt is deterministic from the graph, the same prompt regenerates the same clean answer — reproducible context.

Pilot results

ThoughtDAG includes a Context Repair Pilot with 4 models and 540 runs. It tested three ways to remove an outdated branch from a merged answer:

• Delete source only: 68/72 fixed
• Recompute downstream: 71/72 fixed
• Remove whole subgraph: 72/72 fixed

The pilot demonstrates that context repair works best when you prune the entire offending subgraph — exactly what ThoughtDAG's graph editing enables.

Desktop app

The app is available as a desktop build (v0.3.19):

• macOS (Apple Silicon): .dmg ≈126MB, signed and notarized
• Windows x64: installer ≈106MB (unsigned; SmartScreen shows a warning)
• Linux x64: .AppImage ≈133MB

All versions and history are on GitHub Releases. The macOS build runs with no terminal; the Windows build requires clicking "More info" then "Run anyway" on SmartScreen.

Who it's for

This is for developers and researchers who use LLMs for long, complex reasoning tasks and need to control exactly what context goes into each prompt — without the pollution of unrelated chat detours.