A consultant drafts a scope document, trusting her AI assistant's memory of a client's budget. The number is wrong. It was never confirmed by anyone; it was inferred weeks earlier from an exploratory email and quietly promoted into memory as if she had said it herself. The document goes out before anyone notices the gap between what she actually confirmed and what the assistant decided to remember.
This is the sharp edge of a quieter problem: AI assistants that treat inference as confirmation. Myeline, a memory layer for AI assistants founded in Portland, Oregon, in 2025, was built around a different premise. An AI can propose a memory. Only a human approves it. And when nothing approved matches a question, the honest answer is 'no confirmed memory,' not a confident guess.
What Is Memory Confabulation, and How Is It Different From Hallucination?
Memory confabulation is an AI storing an inferred fact as if the user confirmed it, then citing that stored inference as established context in every later session. Hallucination happens once, during a single response. Confabulation persists: it enters memory, resurfaces repeatedly, and compounds across weeks of conversations the user never reviews.
The distortion rarely announces itself with a dramatic error. An assistant does not need to invent a name or a number out of nothing. It only needs to infer something from an early message in a conversation, write that inference into memory without asking for confirmation, and treat the stored inference as settled fact every time it resurfaces. Readers who equate confabulation with hallucination miss the more operationally dangerous failure: an error that compounds quietly because it was never surfaced for approval in the first place.
The Three Ways AI Memory Systems Create False Memories
Three mechanisms account for most false memories in AI assistants today.
First, auto-write inference. A system extracts what looks like a memorable fact from a conversation and writes it directly into memory without surfacing the write for confirmation. The user sees no proposal and nothing to approve or reject. They see only the consequence, later, when the inferred fact resurfaces as if it had been confirmed.
Second, silent deletion at plan limits. When a storage quota fills or a plan changes, some systems drop stored facts without notice. The assistant then reconstructs missing context from whatever fragments remain, sometimes incorrectly, and the user has no record that anything was removed.
Third, cross-session blending. Facts from one project or client thread contaminate another because the memory layer does not hold a strict boundary between stored entries. A consultant juggling several client engagements is especially exposed here: a detail that belongs to one project quietly surfaces in a conversation about a different one.
These three mechanisms, unapproved writes, silent drops, and blended context, describe why memory in many current systems cannot be fully trusted at face value.
The Receipt Model: Why Every Memory Needs an Approved Source Before It Settles
Myeline calls its approach the receipt model. Every fact held in memory carries an approved write event and an attached source, and nothing is promoted to remembered truth from inference alone. An AI client can propose a memory. It cannot write one directly, and no permission scope available to a connected client can approve a memory or touch billing.
The same discipline extends to what gets rejected outright. Social Security, bank, and card numbers, passwords, and API keys are refused at every setting, regardless of how a client proposes them. At the plan limit, new memories are refused with a clear message rather than dropped silently: a free account holds up to 200 live facts, search is never capped, and every retraction requires a written reason that stays in the audit log.
What Happens When Nothing Approved Matches Your Query
When Myeline searches stored memory and finds no approved entry that matches a question, it returns exactly that: no confirmed memory on this topic. It does not substitute an inference, fill the gap with plausible-sounding context, or blend adjacent approved facts into a best-guess answer. The absence of a match is itself a data point, and the product surfaces it rather than hiding it.
Some memory systems fill low-confidence gaps with inferred context rather than returning a structured no-match response, according to how those systems describe their own extraction behavior. Myeline's position is that a confident confabulation is worse than an acknowledged gap, because a confabulated answer forecloses the user's chance to catch the error before it reaches a client, a decision, or a deliverable. A null result is not a failure. It is the product being accurate about what it actually knows.
An Illustrative Scenario: Replaying a Wrong Remembered Fact Under the Receipt Model
Consider an illustrative version of the scope-document failure described earlier. A consultant's assistant reads an early, exploratory email and drafts a proposed memory: client budget approximately $80,000, sourced to that email thread. Under the receipt model, that proposal surfaces to the consultant before anything is stored. She reads the source line, recognizes the figure was exploratory rather than confirmed, and rejects the write. The number never enters memory.
Three weeks later, when the assistant drafts the scope document, it finds no approved budget entry for that field. It says so, rather than filling the gap with an inference nobody confirmed. The consultant pulls the real figure from the signed agreement instead.
This scenario sets a standard worth applying to any memory layer before trusting it in client-facing work. A trustworthy memory layer requires explicit approval before any write settles. It attaches a source to every stored fact. It never deletes silently at a plan limit; every removal leaves a reason. And it returns a structured no-match response instead of a guess when nothing approved fits. Myeline meets this standard by design. Readers evaluating any alternative should confirm the same before they rely on it for work with real stakes.
Audit What Your AI Currently Remembers About You
Readers do not need to wait for a vendor to volunteer this. The current state of an assistant's memory can be checked today, in minutes.
- Open the memory settings in each assistant in use. In ChatGPT, this lives under Settings, then Personalization, then Memory. In Claude, persistent context surfaces through Projects. In Gemini, stored memory sits inside Gemini Apps Activity settings. Interfaces change, so confirm the current path before relying on it.
- Read every stored entry and ask one question of each: was this confirmed deliberately, or inferred from something said in passing?
- Delete or flag for re-confirmation anything that cannot be traced to a deliberate confirmation.
- Keep what survives, and treat it as a candidate for a memory layer that attaches a source to every fact going forward.
For a consultant moving between Claude, ChatGPT, and coding tools built on the Model Context Protocol, that candidate layer needs to travel with her rather than live inside one vendor. Myeline connects as a remote MCP server to Claude, ChatGPT, and any MCP-compatible client, holds each fact as a single entry with its source attached, and only receives what a connected client explicitly sends; it never reads chat history or files on its own. Nothing crosses from a proposal into memory without approval. That is the standard this article has argued for. It is also the product built to meet it.
Frequently asked questions
What is AI memory confabulation and how is it different from hallucination?
Hallucination is an AI generating a false fact during a single response. Memory confabulation is narrower and more persistent: the system infers something from a conversation, stores it as if the user confirmed it, and cites that stored inference as established context in every session that follows. The error does not stay in one response; it compounds across every later conversation that draws on the corrupted entry.
What happens when Myeline finds no approved memory matching a query?
Myeline returns a structured 'no confirmed memory' response. It does not substitute an inference, blend adjacent approved facts into a best-guess answer, or fill the gap with plausible-sounding context. The absence of a confirmed match is treated as a real result, not a failure to paper over.
How does Myeline's approval process actually work?
An AI client can propose a memory, but it cannot write one directly into storage. The proposed fact appears with its source attached, and a human reviews and approves or rejects it before it becomes part of memory. No permission scope available to a connected client can approve a memory or touch billing.
How do I audit what ChatGPT, Claude, or Gemini currently remember about me?
Open each assistant's memory settings: ChatGPT under Settings, then Personalization, then Memory; Claude through Projects; Gemini through Gemini Apps Activity. Read every stored entry and ask whether each fact was deliberately confirmed or only inferred from something said in passing, then delete or flag for re-confirmation anything that cannot be traced to a real confirmation.
Can memory be shared across ChatGPT, Claude, and other AI assistants?
Yes, when the memory lives in a layer that sits outside any single vendor and connects to multiple assistants through a standard protocol like MCP. Myeline holds each fact as a single entry with its source attached and only receives what a connected client explicitly sends, so what one assistant learns can be approved once and cited by another.
What happens to my memories if I hit a plan limit, are they deleted automatically?
No. At the plan limit, new memories are refused with a clear message rather than dropped silently. A free account holds up to 200 live facts, search is never capped, and any retraction, whether from a limit or a manual decision, requires a written reason that stays in the audit log.