Why does my AI think something about me that I never said?

Usually because the assistant said it, not you. Memory extraction reads a whole conversation and pulls out statements about the user — and an assistant's guess ('so you're clearly a morning person') sits in the transcript looking exactly like a fact you volunteered. Without a record of who spoke, an inference becomes a belief.

What is provenance in AI memory?

Provenance is the answer to four questions about any stored fact: who said it, in what artifact, on what date, and what kind of claim it was — asserted, planned, predicted, reported, or inferred. A memory with provenance can be audited. A memory without it can only be believed or deleted.

Isn't an editable memory list enough?

It's a big improvement and worth using. But an editable list shows you the conclusion, not the evidence. When you read 'prefers formal writing' and don't recognise it, you're left guessing whether it came from one work email in March or from a hundred consistent samples — and those deserve very different treatment.

What's the difference between a plan and an accomplishment?

Enormous, and most memory systems flatten it. 'I'm going to run a marathon in the spring' and 'I ran a marathon' are one tense apart and produce very different assistants. The same collapse happens with opinions read as facts, predictions read as outcomes, and other people's claims read as your own.

Does repetition make a memory more reliable?

Only if the repetitions are independent. Ten copies of the same forwarded email is one source, not ten. Corroboration should require a different artifact, in your own voice — otherwise a single confident mistake reinforces itself every time it's re-read.

How should an AI handle something it isn't sure about?

By saying so. 'You've mentioned X, though I only have one source for it' is more useful than a confident assertion, and far more useful than silence. A memory system that can't express uncertainty will eventually be confidently wrong about something that matters.

Field note · August 26, 2026 · ~7 min read

Did you actually say that?

Open any AI memory list and you will find at least one line you don't recognise. The reflex is to assume the model made it up. Usually it didn't. Usually the model said it — to you, weeks ago, as a guess — and the memory system read the transcript, found a sentence about you, and filed it as something you'd told it.

The four questions

A memory worth trusting can answer four things about itself:

Who said it? You, the assistant, or a third party quoted in the material.
Where? Which conversation, journal entry or document, and on what date.
What kind of claim is it? Something you stated flatly, something you were only considering, or something someone else said about you — three very different things that read identically once they're a bullet in a list.
How well corroborated? One mention, or the same thing in your own voice across several different artifacts.

Answer those and a wrong memory becomes debuggable. Skip them and the only tool you have is deletion, which means fixing your assistant one symptom at a time forever.

Four ways memory goes quietly wrong

Speaker collapse. The assistant's interpretation becomes your fact. This is the single most common cause of "it thinks I'm someone I'm not," and it compounds: the wrong belief shapes the next answer, which produces a transcript that corroborates the wrong belief.

Tense collapse. A plan becomes an accomplishment. "I'm thinking about going back to school" comes back six months later as "since you went back to school." One tense, an entirely different person.

Stance collapse. An opinion becomes a world fact; a prediction becomes an outcome; something someone else said about you becomes something you said about yourself.

False corroboration. The same document imported twice, or a thread quoted in three places, reads as three independent confirmations. Confidence rises with no new evidence.

Why editable lists don't close this

Anthropic's August 2026 update gave people a readable, editable list of what Claude remembers, and it is the best consumer memory control shipped so far. But an entry in a list is a conclusion with the evidence stripped off. When you read a line you don't recognise, you still can't tell whether it came from one throwaway remark or a decade of consistent behaviour — and you'd treat those very differently.

The fix isn't a better list. It's keeping the link between every conclusion and the material it came from, so you can click a belief and read the sentence that caused it.

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 to do about it now

Audit the list monthly. Delete anything you can't source, rather than leaving it because it sounds roughly right.
When you correct something, correct it as a rule ("never assume I've moved") rather than a one-off deletion — patterns outrank single facts in most systems.
Be suspicious of flattering entries. Extraction is biased toward decisive-sounding statements, which are disproportionately the assistant's phrasing, not yours.
Keep the originals. Provenance is only recoverable if the source material still exists somewhere you control.

Frequently Asked Questions

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