CHAT HIDES CONTEXT.
THE GRAPH IS THE CONTEXT.
Linear conversation
Editable context graph
Same prompt · different context
AI conversation 87 messages
Compare three research paths.
Start with the first. Its advantage is…
What if the core hypothesis fails?
Consider another explanation…
Also, what should I eat tonight?
There is a new restaurant nearby.
The history is here. Which parts enter the next request?
research-paper.pdfp.7
Results
The effect appears only in the experimental condition.
Selected from the page
Clipped passageresearch-paper.pdf · p.7
The effect appears only in the experimental condition
Source linked · not wired yet
Asked from sourceresearch-paper.pdf · p.7
What does this evidence actually mean?
The source is in context
Unrelated branchdetour
What should I eat tonight?
This history should not enter the research summary.
Still connected
Polluted summary 3 sources
Research summary… also, consider hot pot for dinner.
The prompt stayed the same. Polluted context changed the answer.
Includes unrelated branch
Will send1,284 tokens
Preview what the model will receive
Incoming ancestors:
Research question Evidence A Dinner detour
After deleting the orange edge: −47 tokens
Context diff−47 tok
The dinner detour left context
Same prompt · regenerate
Same promptask again
Give me a bullet-point summary
The words are identical. Only one edge changed.
Reproducible context
Clean answer 2 sources
One: record the database version. Two: use independent reviewers. Three: resolve conflicts with a third reviewer.
The unrelated dinner suggestion is gone.
Answer updated in place
One ruleThoughtDAG
Wires are context.
No hidden memory selector. What the model sees, why, and what was removed stay visible in the graph.
Visible Editable Inspectable
01 · The problem
Chat history is long. Context is still invisible.
The interface shows what was said, not which history enters the next request.
02 · Externalize
Ask from the source. Clip what matters.
Ask from a selected passage, or turn a passage or figure into its own source-linked node. Provenance stays attached; context remains yours to wire.
03 · Inspect
Before sending, inspect what the model will read.
Preview source nodes, order, and token count. Context is no longer a hidden decision.
04 · Edit
Delete one edge. Ask the same question again.
The removed branch really leaves the request. The answer changes with the context.
05 · The protocol
Most canvases organize information. ThoughtDAG edits context.
You decide what enters and leaves. The graph is the context protocol before generation.
1 / 5 Invisible context
Instagram Engineering
Shopify Engineering