Using Attractor to ask what conversations return to, and how the models reading them shape the answer.

Jennifer Naomi Nguyen · research tooling · Exploratory work · March 2026 – present · built with Claude Code and Codex

README · ARCHITECTURE · TECHNICAL · source

The question that got me here

When I talk with a model over many sessions, the conversations seem to return to familiar topics and ways of thinking. I want to understand how those patterns develop — and whether supplying memory helps create the patterns the system later reports.

This uses the same mechanism as Attractor — Adaptive Context and Personalization. That project asks what context to deliver. Here, I'm using the tool as an instrument for research.

What I can explore with it

Attractor reads a conversation, summarizes it, and asks a model how it would update a weighted map of engagement modes. Its compare command runs the same conversation and starting state through several model and engine choices, then puts the proposed updates side by side.

I can look at the summaries, weight changes, connections, and emerging patterns. The local run log groups outputs by conversation and model, which makes it easier to return to an interesting difference.

A small example

In one recorded conversation with two runs per model, Haiku proposed +0.15 twice for the main Systems-design basin. Opus proposed +0.09 and +0.08, and returned three or four emerging patterns where Haiku returned none.

That's enough to give me a more specific question: does that difference persist when I hold the summary fixed and repeat the comparison across more conversations? The current comparison includes each model's own summary, so the technical notes unpack what is being compared.

What you can inspect

I gave Luna and Astra the same Journal Club README and four topic basins, each starting at 0.30. Both put research first, but their proposed connections differed: Astra also linked research to routing and infrastructure, pointing to how response-parsing problems affected the study. Neither proposed a direct routing-to-memory connection in this run.

The interactive demo lets you switch between the saved maps and inspect the keywords, weights, and connections.

This is one saved run per model using only the README, fixed categories, and no new-basin creation. The maps show proposed relationships rather than code structure, and opening them makes no model calls. The technical notes explain how to read the graph and what the comparison does not establish.

Where the research goes next

• Look outside the instrument. Compare Attractor's map with patterns measured from the conversations themselves.
• Repeat the model comparison. Use fixed summaries and starting states so the update stage is easier to study.
• Vary the memory supplied. Compare conversations with and without the context block to investigate feedback.

The conversation-pattern study lays out those questions and the planned experiments. The tool and preliminary examples exist; the controlled studies are still ahead.

How I'm approaching it

The interesting part is the connection between building a tool and learning what its output means. Labels, summaries, update rules, and model choice all shape the map. Making those parts visible gives me something concrete to investigate.

The source section connects the implementation and comparison examples.

All projects · Attractor — Adaptive Context and Personalization