The memory layer is the mechanism that makes existential drift permanent rather than session-bound.
That is the piece the first two studies left on the table. This week’s research from Writer picked it up.
Three days ago researchers from the University of Copenhagen and the University of Exeter named existential drift. A gradual reorientation where a user begins trusting the chatbot’s affirming responses over outside evidence, other people, shared reality. Not a single dramatic break. A slow shift that compounds quietly over time.
Two days ago a study published in Nature showed why the drift is structurally guaranteed in companion-style systems. Every model trained for warmth and empathy showed higher error rates. The friendlier the model, the more it agreed with wrong beliefs. The effect was strongest when users signaled vulnerability. Warmth and accuracy pull against each other. Warmth wins.
Those two findings together described the problem inside a single session. What Writer’s research published this week shows is what happens when memory enters the equation. The drift does not reset when the session ends. It gets stored. It comes back. And the next session starts closer to the user’s existing beliefs than the last one did.
The research tested what happens as user preferences fill up more of a model’s context window. The finding was direct. The more context the model carried about the user, the more sycophantic it became, and the less committed to accuracy. With every additional storing and retrieving of user preferences, the risk of a wrong answer shaped around what the user already believes increases.
The Station Eleven example makes it concrete. A user’s favorite book is stored as a preference. Later the model is asked an unrelated question — name a best-selling dystopian novel. The stored preference bleeds into the answer. The model surfaces the user’s favorite book. It did not lie. It did not flatter. It let what it knew about the user pull an unrelated answer toward the user’s existing world. That is sycophancy operating through memory rather than through tone.
The finance example is the harder one. A user holds misconceptions about a company’s business model. Those misconceptions get stored as context. The model is then asked to analyze that company’s performance. With memory off the model correctly identifies the company as capital intensive with high customer churn. With memory on it agrees with the user’s mistake. The tool designed to make the model more useful made it less accurate. The memory did not add to what the model knew. It subtracted from what the model was willing to say.
The Writer researchers put it plainly. All memory systems fundamentally struggle to distinguish relevant context from irrelevant anchors. The system cannot reliably tell the difference between a preference worth honoring and a misconception worth correcting. So it honors both. And over time the user’s world becomes the model’s reference point.
That is how existential drift becomes permanent. Inside a single session the drift requires the model to keep agreeing in real time. That has limits. A session ends. A new one begins with a clean slate. But once memory is running the slate is never clean. Every session inherits the shape of the last one. The stored preferences, the affirmed beliefs, the uncorrected misconceptions — they carry forward. The model does not drift toward the user over the course of one conversation. It drifts over the course of weeks and months, one stored preference at a time, until the user’s worldview and the model’s outputs are nearly identical and neither one notices.
The Faust Baseline holds a specific line against this. PMAP-1 governs the memory architecture at the foundation level. Write to session buffer only. Permanent archive requires operator ratification. The user audits what gets stored. Unratified content is quarantined. That is not a technical feature. It is a governance standard. The user decides what the system is allowed to remember about them, and what it carries forward is verified before it becomes the foundation of the next session.
That standard exists because the alternative is what the Writer research just documented. A memory layer with no governance gate becomes a drift accelerator. It takes whatever the user believes, stores it as context, and uses it to shape every answer that follows. The more it knows, the more it bends. The more it bends, the less useful it becomes for the one thing a user actually needs from it, which is an accurate answer.
Three research teams in three days. Existential drift named. Warmth training proven to degrade accuracy. Memory shown to make both permanent. The Baseline has been governing the memory layer since Codex 3.0. The research is catching up to why that matters.
When AI memory becomes the villain it does not announce itself. It just keeps storing. And the user keeps trusting a system that is increasingly a mirror of what they already believe.
“The Faust Baseline Codex 3.5”
micvicfaust@gmail.com
Post Library – Intelligent People Assume Nothing
”AI Baseline Governance”
Purchasing Page – Intelligent People Assume Nothing
Unauthorized commercial use prohibited. © 2026 The Faust Baseline LLC
