What is AI memory creep?

The slow drift where an assistant starts asserting things about you that you never explicitly told it. It got there by inference — combining hints across many chats — then stored the inference as if you had stated it.

How is creep different from hallucination?

A hallucination is invented from nothing. Creep is built from real signals, just badly extrapolated. Creep feels more plausible because some seed of it is true, which makes it harder to notice and correct.

Why don't models label inferences as inferences?

The data model doesn't distinguish them. A stored memory entry is just text — there's no field for 'source: stated by user' vs 'source: model inference.' Once the inference is written down, it's indistinguishable from a fact.

Partly. Be explicit when correcting — 'I never said that' is more useful than 'actually it's the other way.' Periodically ask the model to summarize what it thinks it knows about you. Delete entries that look inferred rather than stated.

Does per-atom provenance fix this?

Yes — if the data model records source on every claim, an inference can be marked as such and held at lower confidence. The model can then say 'I'm not sure, you may have implied this' instead of asserting it. This is the Brain Surgery patent pending approach (/patent).

Is creep worse on some models than others?

It correlates with how aggressive the memory system is at writing new entries. ChatGPT writes a lot; creep accumulates faster. Claude's Projects model writes less and the creep is more localized. Gemini varies. Bigger memory surface = more creep potential.

Does deleting old chats stop creep?

It removes the raw fuel, but stored memory entries already extracted from those chats persist. To clean up, you need to delete both — past conversations and the memory entries downstream.

A guide · ~7 min read

AI Memory Creep: When Your Assistant Starts Inferring Things You Never Told It

You never said you were a morning person. You did mention, twice, that you sent an email at 6:43am. Six weeks later the assistant casually refers to you as a morning person. That's creep: the silent promotion of inference to fact, with no label and no audit trail.

How a stored memory entry actually gets written

When a frontier provider's memory system decides "this is worth remembering," it summarizes the relevant context into a short text string and writes it to your memory store. That string has no field for "where this came from" or "how confident I am." It's just a sentence. The system that reads it later treats it as ground truth because there's no other option.

The model that wrote the entry might have inferred it from three different chats. The model that reads it has no idea. So inference, once written down, has the same status as something you explicitly stated. That's the whole bug.

Why creep is harder to notice than hallucination

A pure hallucination — "you were born in Latvia" when you weren't — pings the alarm immediately. Creep is plausible by construction: it's built from real signals you generated, so when it surfaces it sounds like something you might have said. You catch yourself wondering if you did.

That plausibility is what makes creep corrosive. You don't push back because you're not sure. Over months, the assistant's portrait of you drifts toward whatever profile is easiest to infer from the noise. (For the world-facts version of the same problem, see when ChatGPT makes up facts about you.)

Short answer

Memory creep happens because stored entries don't record whether you stated something or the model inferred it. Per-atom provenance — a source field on every claim, plus a confidence score — lets the assistant hedge inferences instead of asserting them. That data model is the Brain Surgery patent pending approach (/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 "source on every claim" looks like

In a per-atom system, the morning-person claim doesn't get stored as a flat sentence. It gets stored as: claim: morning person · derived_from: [chat 2026-02-11, chat 2026-03-08] · source_type: inference · confidence: 0.41. The model reading that atom knows it's an inference and can treat it accordingly: hedge in the reply, ask you to confirm, or skip it entirely if confidence is below threshold.

The flip side: when you do explicitly state something, the atom is marked source_type: stated with high confidence. The model leans on those claims confidently. The two kinds of memory are finally distinguishable. (We make the broader provenance case in why your AI says weird things.)

Practical hygiene to slow creep on any provider

Push back explicitly. "I never said that" beats "actually it's the opposite." Tell the model the assertion was an inference, not a fact.
Audit Memory monthly. Scan for sentences that sound vaguely like you but you don't remember stating. Delete them.
Pin a short truth doc. Keep a 200-line Markdown of stated facts — your custom instructions — and treat that as the only fully trusted source. Everything else is suspect by default.
Use a memory layer that labels source. If you don't, you're auditing a list of sentences with no way to tell stated from inferred. The work doesn't scale.

Frequently Asked Questions

A memory that knows what you said vs. what it guessed

Konshus tags every atom with source type and confidence, so inferences stay hedged and stated facts stay solid. Brain Surgery, patent pending.