When an AI Remembers Something Wrong About a Client
Independent consultants running Claude, ChatGPT, and two or three other assistants across client work know this moment. An assistant states a client's reporting cadence as settled fact, except it is wrong, and the error has already worked its way into a deliverable before anyone catches it. Nobody approved that fact. Nobody saw it get stored. The assistant simply decided, at some earlier point, that it was worth remembering, and wrote it down without asking.
This is the scenario that makes AI memory a trust question rather than a convenience question. The anxiety is not about whether an assistant can recall a preference quickly. It is about whether the system holding that preference can be trusted to hold it correctly, and whether anyone would know if it couldn't.
Myeline is built on a different premise: memory is not something an AI should decide on its own. Every fact a connected assistant wants to keep is proposed, sourced, and held for approval before it becomes part of the record.
Why AI Memory Is a Trust Problem Before It's a Productivity Problem
Most AI memory products are marketed on speed: faster recall, less repetition, no re-explaining yourself to a new chat window. That framing hides an assumption. It assumes the AI's judgment about what to remember, and what to discard, is good enough to act on without checking.
That assumption is the actual product being sold, even when the marketing talks about convenience. A system built on automatic or continuous extraction watches a conversation, infers what matters, and writes it to memory. The user's first contact with that decision often comes later, when the assistant repeats something back that was never verified.
Myeline starts from the opposite assumption. A connected AI client can propose a memory. It cannot write one. The decision about what becomes a permanent fact about a client, a project, or a preference stays with the person who has to live with being wrong about it.
What Is Approval-Gated Memory? Staged Writes vs. Auto-Extraction
Approval-gated memory is a memory architecture in which an AI assistant can propose a fact for storage but cannot commit it. It works by staging every candidate fact with its source, holding it in a review queue, and writing it to permanent memory only after a human approves it.
Myeline calls the alternative a staged write, and its documentation at myeline.si/docs draws the contrast directly. Auto-extraction treats inference as permission: when an assistant decides a fact is worth keeping and writes it without asking, the moment of decision is invisible, and the audit trail is already gone. A staged write does not guess at intent. It surfaces the proposed fact, names its source, and waits. The write happens only when you say so.
Tools that describe themselves as using automatic or continuous memory extraction position that behavior, in their own public documentation, as a way to reduce friction. The tradeoff they do not advertise is that reduced friction means reduced visibility: the write already happened by the time a user could object to it. Approval-gated memory keeps that friction, on purpose, at the one point where it protects the user: before anything becomes a fact.
How Does Myeline's Staged-Approval Queue Actually Work?
Myeline's staged-approval queue works by holding every AI-proposed fact in a reviewable state until a person acts on it. No fact enters active memory, and no connected assistant can act on it, until that approval step is completed, whether the fact came from a chat conversation or a coding agent's context.
- The connected AI assistant identifies a candidate fact during a conversation or a document it is working from.
- A proposed-fact card appears in Myeline's review queue, showing the candidate fact next to the exact source it came from.
- The user reads the proposed fact and its cited source.
- The user approves or rejects the card.
- On approval, the fact is committed to memory with its source attached, and it becomes available to every connected assistant.
- On rejection, the event is logged with a timestamp and a written reason, which stays in the audit log.
No write happens before step four. That is the entire mechanism, whether the proposing client is a chat assistant or a coding agent connected over MCP.
What Myeline's Permission Scopes Can and Cannot Do
Myeline's public documentation describes three permission scopes for connected clients: cortex:query (read), cortex:propose (save for review), and cortex:retract (retire facts). Not one of them can approve a memory, and none can touch billing or account configuration. Approval exists at no scope at all: the only place a fact becomes live is the workspace, under the signed-in person's own click.
This is not a settings toggle an administrator could switch on. There is no approve scope to grant, no token that carries it, and no API path that accepts it: a misconfigured integration, a compromised token, or an overly broad connection cannot promote itself into approval rights, because there is no access level at which that promotion exists.
For a technical founder evaluating connectors before standardizing a team on MCP, this is the detail worth verifying directly: least-privilege is not a policy Myeline asks you to trust. It is a constraint enforced in the architecture itself, documented at myeline.si/docs.
Every Fact Has a Source: Provenance and Retraction in Practice
Every memory Myeline stores is a single fact with its source attached. When a consultant's assistant proposes that a client prefers weekly reporting, the approved memory carries a citation back to the specific conversation where that preference was stated, not just the claim itself.
If that source conversation later turns out to be wrong, the correction is not silent. Forgetting a fact in Myeline takes a written reason, and that reason stays in the audit log even after the fact stops appearing in answers. A consultant who needs to reconstruct what an assistant knew, and when, at a given point in a client engagement can do that from the log.
The same discipline applies when nothing matches. If a search returns no approved fact, Myeline says so rather than guessing. And at the free plan's limit of 200 live facts, a new memory is refused with a clear message rather than quietly dropped. Myeline also refuses to store Social Security numbers, bank and card numbers, passwords, and API keys, at every plan tier, before anything is saved. Source attribution, honest refusal, and a persistent audit trail are the same design decision applied in three places: nothing enters or leaves memory without a visible record.
One Memory Layer Across Claude, ChatGPT, and Every MCP Client
A consultant running three assistants in parallel usually ends up with three separate, unaudited versions of the same facts, because each tool only remembers what happened inside its own conversations. The tools drift apart. They contradict each other because none of them learned from the others.
Myeline connects to Claude, ChatGPT, and any client that speaks the Model Context Protocol, and it only receives what those clients explicitly send through its tools. It never reads chat history, a client's own built-in memory, or files on its own. A fact approved once in Myeline's review queue becomes available to every connected assistant, with its source and approval record intact.
For a developer connecting Claude Code, Cursor, or a custom agent runtime, that single approved-only layer prevents a specific failure mode that tools built on automatic extraction cannot rule out: an agent acting on a memory nobody reviewed. If the memory was never approved, it was never written, so there is nothing polluted for the agent to inherit.
What to Do Next: Review What Your AI Already Remembers
Every assistant a reader uses today, whether or not it is connected to Myeline, is already forming opinions about what to remember. The first useful step is finding out what those opinions are.
For a consultant juggling client work across tools, that means connecting Myeline to each assistant and letting proposed facts collect in the review queue rather than approving anything automatically. For a technical founder preparing for a security or customer review, it means asking any memory vendor the same three questions this article answered: what can write to memory, what can an admin approve without the user, and what happens when nothing matches. For a developer building on MCP, it means testing the staged-approval queue against a real coding session before trusting an agent's memory with production decisions.
Myeline starts free, with room for 200 live facts, so the first review queue costs nothing to open.
Frequently asked questions
What is approval-gated memory?
Approval-gated memory is a memory architecture in which an AI assistant can propose a fact for storage but cannot commit it. It works by staging the candidate fact with its source, holding it in a review queue, and writing it to permanent memory only after a human approves it. This is the opposite of auto-extraction, where an assistant infers and writes a fact without asking.
What happens if I reject a proposed fact in Myeline?
The rejection is logged with a timestamp and a written reason, and the fact is not written to memory. Nothing is silently discarded: the log documents what was proposed, what you decided, and when.
Can an admin approve a memory write on my behalf?
No. Myeline's three permission scopes are cortex:query (read), cortex:propose (save for review), and cortex:retract (retire facts), and none of them can approve a memory. There is no admin scope and no approve scope at all: the write pathway does not exist at the API layer, so no scope can be promoted into it. Approval happens only in the workspace, under the signed-in person's own click.
How does Myeline share memory across Claude, ChatGPT, and other AI tools?
Myeline connects to Claude, ChatGPT, and any client that speaks the Model Context Protocol, and it only receives what those clients explicitly send through its tools. A fact approved once in the review queue becomes available to every connected assistant, with its source and approval record intact, so context does not need to be rebuilt in each tool.
What happens when a stored fact is deleted or retracted?
Forgetting a fact in Myeline takes a written reason, and that reason stays in the audit log even after the fact stops appearing in answers. The retraction is not silent: a user can later reconstruct what the system knew, and when, at a given point in time.
What does Myeline refuse to store?
Myeline refuses to store Social Security numbers, bank and card numbers, passwords, and API keys, at every plan tier, before anything is saved. It also never reads chat history, a client's own memory, or files directly; it only receives what connected clients explicitly send through its tools.