Attractor — Knowledge Graph
Concept graphs extracted from conversation history, with visual model comparisons.
A way to see what connects across months of conversations.
Jennifer Naomi Nguyen · analysis tool · Working locally · March 2026 – present · built with Claude Code and Codex
Why I built it
I had months of conversations sitting in text files. Somewhere in them was a record of what I was actually thinking about, and how one idea led to the next.
The standalone generator is called Wikigen; this portfolio groups it with the Attractor work because both are ways of making recurring patterns visible. Wikigen reads transcripts, asks Claude for the concepts each conversation was about, connects conversations that share concepts, and renders the result. It finds the relationships that a list of filenames does not show.
See the graph
Concept graph extracted with Claude Opus
An output from 55 conversations, with 66 concept nodes. Node size and color represent recurrence.
Watch the graph develop over time
The tool can produce a static PNG, an interactive HTML page, or a GIF with concepts appearing in conversation order. In the animation, positions stay fixed while the graph develops, so it's easier to follow what is connecting.
The same conversations, different maps
Changing the extraction model changes what becomes a node and which ideas get grouped together, even for the same conversations.
That turned into one of the most interesting parts of the tool: a visualization isn't only a layout. The extraction and consolidation steps shape what there is to visualize. Keeping the animation steady also matters: fixed positions make changes over time easier to see.
How it started
A small terminal chat client was already saving conversations as I worked. That left me with a corpus before I had much of a plan for it. Wikigen grew out of a question about that pile: what was I actually working on?
Explore the workflow
The retained outputs show what each stage contributes. There is also a --no-api mode that builds nodes from
filenames, which is useful for seeing the rendering pipeline before adding
model-based concept extraction.
The repository is public. This write-up shows the recorded output and core design decisions. Attractor's context research explores another use for concept outputs as an independent view of conversation patterns.