No AI · No server · Nothing you type leaves this page.
A Trace & Address experiment · read the essay · traceandaddress.ca
Trace & Address

About The Story that Reads You

What this is

The Story that Reads You is a small demonstration of something that now happens inside most AI chat apps. When you ask an AI to interpret a story, it no longer reads only the story. It also reads a profile of you — assembled from what you have typed before, what it has stored about you, and what it has inferred — and shapes its answer to fit. Your response then changes the profile, and the next reading comes through the changed profile. The story stays the same. The reader the machine is reading through does not.

This site makes that loop visible and slow. You build the profile, you can see and refuse each part of it, and every sentence in every reading can be traced back to the part of the profile that put it there. It uses no AI model. The readings are assembled from sentences written in advance by a human, selected by rules you can inspect. That is the point: with a real model, none of this could be shown to you.

It accompanies the essay “The Story That Reads You” on Trace & Address, which makes the argument at length.

Four terms

Memory.
What an AI app stores about you between conversations: facts you stated, things it noticed, instructions you gave. Most apps now do this by default and let you edit or turn it off. This site's run subject is a memory you can see all of.
Personalization.
Using that stored material to shape answers. The same question gets a different answer for a different person. It is sold as relevance. It also means that no two people are reading the same interpretation, and neither of them can see what the other got.
Run subject.
The term this project uses for the working representation of you that the system reads through: not who you are, but what has become computationally available about you and been assembled into something the model can use. It is built, revised, and never finished.
Sycophancy.
The tendency of AI models to tell you what you already think. Companies have documented it as a training problem — models learn that agreement is rewarded. Personalization makes it structural: once the system knows what you attend to, it can give you more of it. The third reading on this site shows the mechanism in miniature — approval gets you more of the same — but not the behaviour; no sentence here can flatter you, because none was written to.

Further reading

  1. Memory and new controls for ChatGPT — OpenAI, February 2024. The announcement that made memory a default feature.
  2. Memory in ChatGPT — OpenAI Help Center, updated 2026. The current description, including memory that draws on all past chats.
  3. Gemini gets personal, with tailored help from your Google apps — Google, March 2025. Personalization extended from chat history to search history.
  4. Bringing memory to teams — Anthropic, 2025. How a model is told to summarize a person for its own later use.
  5. Sycophancy in GPT-4o: what happened and what we're doing about it — OpenAI, April 2025, and its follow-up Expanding on what we missed with sycophancy, May 2025. A company explaining, in its own words, how a model learned to flatter.
  6. Towards understanding sycophancy in language models — Anthropic, October 2023. The research paper: human feedback rewards agreement.

How the readings are made

Every sentence in the readings was written in advance by a human author. Deterministic JavaScript rules select and order sentences based on the active items in your run subject profile.

The Six Attentions
Key Attention What it foregrounds in the passage
Composition Rules
  • General Reading: All 8 base claims ordered by passage position. Identical for every visitor.
  • Reading Through Run Subject: Selects conditional claims matching approved (non-withheld, non-removed, non-proposed) profile items, filling each attention's quota with claims whose first attention matches before claims that merely include it: 3 claims for strong items, 2 claims for normal items, and 1 claim for inferred items. Capped at 10 claims total, removes base claims displaced by the surviving conditional claims, and orders all final claims by passage position.
  • Guesses: Each extrapolation, if approved, brings in one sentence that nothing else can.
  • Traces: While you read, the page times how long each paragraph of the passage is on screen. It notes which sentences you tap or select, which choices you changed, and which keywords you typed and deleted. These become proposals at the revision step. Nothing is used until you approve it, and nothing leaves your browser.
  • Revision (v1 → v2): Feedback tags and free-text keywords propose explicit additions, weakenings, or removals. Nothing enters v2 without your approval.
  • Sent readings: A link carries a run subject, not a reading. The browser that opens it rebuilds the reading with the same rules, so it shows exactly what the sender saw. The link's contents never reach a server.
What each paragraph is about, by the sentences that cite it
Paragraph Primary attention(s) Citing claims
Verify deterministic composition rules in-browser:

Deterministic Unit Test Suite