What's a hallucination about me, specifically?

It's when the model produces a confident-sounding claim about you personally — your job, location, family, beliefs, projects — that you never told it and that isn't in any source it should have access to. World-knowledge hallucinations get most of the press, but person-specific hallucinations are what break trust day-to-day.

Why is this worse than a normal hallucination?

Because you can't fact-check it against the web. If ChatGPT misquotes a Supreme Court case you can look it up. If it claims you 'usually work weekends,' the only ground truth is your own memory — and you start to wonder if you said something you don't remember saying. That's a corrosive failure mode.

Why do models do this when they 'know' they shouldn't?

Large language models don't have a clean separation between retrieval and generation. When asked about you they blend stored memory entries, in-context messages, and statistical priors about people who sound like you. Without a source-pinned data model, the output is a soft average — confident in tone, fuzzy in origin.

Does turning Memory off prevent person-specific hallucinations?

It reduces the surface area but doesn't eliminate the problem. Even in a single session, the model will infer facts from context and present them confidently. The deeper fix is to give the model a memory layer where every claim has a source and a confidence score, so the model can say 'I don't actually know that' instead of guessing.

How does source + confidence per atom help?

Every fact stored about you is one atom: the claim, the source artifact it came from, and a confidence score. When the model is asked something, it can only assert claims it has atoms for, and it can flag low-confidence ones. Hallucinations become refusals or hedges, which is exactly what you want. This per-atom provenance is the Brain Surgery patent pending claim (/patent).

Can I correct a person-specific hallucination once it's happened?

In native ChatGPT/Claude/Gemini, you can ask the model to forget, edit the Memory panel, or open a new chat. None of these guarantee the bad claim doesn't resurface, because the underlying inference pattern is still there. In a per-atom system, you can pin a corrective atom with high confidence that overrides the bad one, and the audit trail shows you it stuck.

Is Konshus immune to hallucinations?

No — we're built on top of frontier models, so the underlying capacity to hallucinate is still there. What we change is the data model: replies are grounded in atoms, atoms carry source and confidence, and you can see and edit both. That doesn't eliminate hallucination; it makes it visible and fixable.

A field guide · ~8 min read

When ChatGPT Makes Up Facts About You: A Field Guide

You know what to do when an AI hallucinates a case citation or a chemical formula — you check it against the world. But what do you do when it hallucinates a fact about you? There's no external ground truth. There's just the slow, unsettling sense that the assistant has built a portrait of you that you didn't authorize.

Why person-specific hallucinations are a different beast

A factual hallucination about the world ("the Treaty of Westphalia was signed in 1652") is annoying but easy to catch. Wikipedia exists. The hallucination about you ("you mentioned you prefer evening meetings") slides past every filter because the only fact-checker is your own memory, and your own memory is fuzzy on whether you said exactly that, in those words, three months ago in some chat you barely remember.

The result is a quiet erosion of trust. You start hedging your assistant's outputs, second-guessing its summaries, and eventually stop relying on it for anything that requires it to remember you accurately. The assistant didn't get worse — you just noticed.

The mechanics: where these claims actually come from

Memory blending

Frontier providers store memory as freeform text snippets. When you ask the model a question, those snippets get pulled into context alongside the current chat and the model's statistical priors. The model produces a single fluent reply that's an average of all those inputs, which means a claim can emerge that wasn't in any single source.

Inference promoted to fact

The model is allowed to infer things from context — that's part of what makes it useful. But it isn't required to label the inferences. Once an inference appears in a reply, the next session may treat it as something you said. We unpack this in detail in AI memory creep.

Cross-user contamination (mostly a myth, partly real)

Your data isn't directly mixed with another user's at inference time — providers do isolate that. But the base model's training data and reinforcement learning shape its priors about "a user like you," and those priors can show up as confident assertions. It's not your neighbor's data; it's the statistical ghost of a million people who sound like you.

Short answer

ChatGPT, Claude, and Gemini hallucinate facts about you because their memory layer stores claims without sources or confidence scores. The structural fix is per-atom provenance: every fact pinned to the artifact it came from, with a confidence value the model can hedge against. That data model is the Brain Surgery patent pending claim (/patent).

Never lose your AI again

Konshus is one way to solve this — a persistent memory vault and portable persona that follows you across ChatGPT, Claude, Gemini, and whatever ships next.

Meet Konshus

What a source-pinned memory looks like in practice

Concretely: instead of a memory entry that says "user prefers evening meetings," you have an atom that says claim: prefers evening meetings · source: chat 2026-03-14 msg 7 · confidence: 0.62. When the model is composing a reply, it can use the atom — and if you ever ask "where did you get that?", the system can show you the exact message that produced it. If you disagree, you delete the atom or write a corrective one with higher confidence. The reply changes from the next turn forward.

That sounds like overhead. In practice it's the only way the assistant can stop guessing. Without sources, every claim is a guess wearing a confident voice; with sources, the model can finally say "I don't actually know."

Four habits that reduce person-specific hallucinations now

Periodically ask "what do you think you know about me?" Save the answer. Compare to last month's. Hallucinations usually surface in the diff.
Treat your custom instructions as the source of truth. Keep a single Markdown doc with the facts that matter, paste it at the start of any chat that needs them, and don't trust stored memory for anything you couldn't reconstruct.
Correct in writing, not in conversation. "Actually I work mornings" in a chat is weaker than editing the offending memory entry directly. Use the Memory panel.
If you care about this, use a memory layer with provenance. See the full why-AI-says-weird-things piece for the longer argument.

Frequently Asked Questions

A memory layer that can say 'I don't actually know'

Konshus stores claims as atoms with source and confidence, so your assistant grounds its replies in things you actually said. Brain Surgery, patent pending — full claim list at /patent.

Meet Konshus