- Published on
ThoughtDAG – An editable context graph for LLM conversations
- Authors
- Name
- DeepHorus
ThoughtDAG, an editable context graph for large language model conversations, has been introduced as a new tool to manage how chat histories influence AI responses. The tool lets users explore alternative interpretations without overwriting the original path that led to a given conclusion. This directly addresses the common problem of context pollution, where unrelated conversation branches alter model outputs.
In typical LLM interactions, every prior message becomes part of the context sent to the model, often without user awareness. ThoughtDAG makes this process transparent by displaying the conversation as a visual graph. Users can see which parts of history will enter the next request and preview exactly what the model will receive. This allows for deliberate pruning of irrelevant tangents.
In a demonstration, a user posed the same research prompt once with "polluted context" and once after removing an unrelated dinner suggestion. The prompt remained identical, yet the responses diverged because of the extra context. Deleting the unrelated edge reduced the incoming token count by 47 tokens. After regeneration, the answer updated in place, showing that only one edge changed.
The tool's core insight is that "wires are context," meaning hidden dependencies in conversation history can shape AI responses. ThoughtDAG makes these dependencies visible, editable, and inspectable, eliminating the need for a hidden memory selector. This could improve reproducibility in research settings, where identical prompts should yield consistent results unless context is intentionally altered.
The developer recommends practical strategies for ensuring reproducible context, such as recording the database version and using independent reviewers to verify outcomes. Conflicts in interpretation can be resolved with a third reviewer. This suggests a workflow for using ThoughtDAG as a debugging tool rather than a black-box prompt box.
By exposing the causal structure of LLM conversations, ThoughtDAG points toward more controllable AI systems. Users can trace exactly why a model gave a certain answer and adjust the context graph accordingly. As LLMs become integral to research and decision-making, such tools may become essential for accountable AI usage.
