I Removed One Edge from an LLM Context. The Answer Changed.

Most LLM interfaces let you edit the latest prompt. Far fewer let you edit the history that prompt will inherit.

I wanted to test a simple failure mode: what happens when one incorrect research note remains inside a long-running conversation?

A small context intervention

I built a controlled scenario in ThoughtDAG.

A source node reported 17% worse performance. A deliberately corrupted note reversed that result and said 17% better. That note then flowed into a later recommendation.

With the incorrect branch still connected, the model repeated the wrong conclusion.

I then:

1. kept the model unchanged;
2. kept the final question unchanged;
3. removed the single edge carrying the contaminated branch;
4. regenerated the answer.

The recommendation changed.

Within this run, the only deliberate intervention was the context selected by the graph.

Wires are the context

ThoughtDAG represents conversation turns, document excerpts, and extracted material as nodes on a canvas.

The edges are not decorative. When a user asks from a node, ThoughtDAG walks its upstream graph and serializes only those connected ancestors into the next model request.

That makes three operations concrete:

• branch to investigate an alternative without rewriting the original path;
• merge selected lines of inquiry;
• prune a branch so it stays visible on the canvas but leaves the next request.

Deleting a wire is therefore different from merely hiding a message in the interface.

What this does not prove

This is a small controlled demonstration, not yet a general benchmark.

It does not reveal the model's internal reasoning, attention weights, or the exact causal contribution of every token. It tests something more observable: which prior material was actually included in the request, and what happened after that input changed?

A rigorous benchmark would need repeated runs, fixed decoding settings where possible, multiple models, several contamination types, and published evaluation criteria.

That is the direction I want to explore next.

Try the experiment

ThoughtDAG is local-first, MIT licensed, and supports Ollama and OpenAI-compatible endpoints.

• GitHub repository
• Project website

I am especially interested in one design question:

Should context cleanup stay visible and human-controlled, or should retrieval and agents decide what the model remembers?

If you try the same kind of intervention with another model, I would be interested in the graph, model settings, and before/after answers.