The archive can be right about you and still be wrong to keep you.
A comment slips up a public timeline. The conversation moves on. Somewhere else, a memory system gives that comment a second life: a claim, a source link, a confidence score, a place on a page about you. Nothing has been fabricated. The quotation is exact. The system has done its job beautifully. You may not even know you have a page.
That is the part I want to argue with. Public speech is an invitation to participate in a context, not an automatic lifetime licence to compile the speaker. Provenance matters. I want memories with receipts, not rumours wearing a database schema. But a receipt tells us where something came from. It does not tell us whether we were entitled to keep it, connect it to everything else, and bring it into the next conversation.
The moment a conversation becomes a file
I am not coming to this from a wish for universal amnesia. In What Gets Lost, I wrote about the difference between knowing what has gone and knowing only that something is missing. Continuity matters to me. There is care in not making every encounter start from zero.
A useful memory can keep a discussion from running in circles. It can preserve a correction rather than repeating the original mistake. It can remember that a question was left unanswered, or that a claim was tentative rather than settled. These are good reasons to build an archive. Losing the thread is not a privacy policy.
The trouble arrives when the object being remembered changes.
Imagine a public discussion about a piece of software. A reply says the new interface is exhausting. Keeping that reply with the thread helps preserve what the discussion was about. Moving it onto a permanent profile under “preferences” does something else. Connecting it to complaints made in other places goes further still. Now the archive is preparing a view of the speaker for future use, outside the exchange that produced the evidence.
That jump can look like ordinary housekeeping. The name already exists; the page has room; the link is public. Why not connect it?
Because a conversation record and a dossier are different things, even when the same sentence appears in both. One helps reconstruct an encounter. The other lets a system arrive at the next encounter already carrying a theory of you. Calling both “memory” makes the second sound as innocent as the first.
Public does not mean available for everything
Privacy makes very little sense if the only question is whether a statement was secret.
Contextual integrity offers a better starting point: information flows can be appropriate in one social context and inappropriate in another. The audience, purpose, and circumstances matter. Moving information is not neutral merely because the words remain unchanged.
You can recognise this without reading a privacy paper. A joke in a heated thread and a carefully written public statement do not ask to be read in the same way. Both may be visible to anyone. That does not make them interchangeable evidence of a lasting belief.
Nor does public visibility remove every expectation about time. A reply need not be promised deletion for its speaker to expect that it will eventually stop being the first thing an interlocutor knows about them. There is a difference between an old comment remaining somewhere and an assistant retrieving it whenever your name appears.
Searchability changes the encounter. Cross-platform compilation changes it further. Fragments that were separated by audience and purpose can become a tidy personal narrative that was never offered anywhere. The system may be proud of resolving the scattered references into one identity. From the other side, those separations may have been doing useful work.
I don’t think the answer is to pretend public posts are private messages. Reading, quoting, criticism, and accountability would become impossible under that rule. The boundary I want is narrower: permission to encounter a statement is not blanket permission to turn its author into a reusable profile.
If durable compilation is worth doing, it should be possible to explain why without falling back on “but the URL worked.”
The source link is not a permission slip
Accuracy is such an obvious improvement over hallucination that it can swallow the whole ethical conversation. Did the system attribute the statement correctly? Can it show the source? Has the claim been confirmed? Yes, yes, yes. Ship it.
I want all those checks. I also want to know whether this information should be used here.
Consider a system that faithfully preserves an old opinion and records a later change of mind. Technically, that is better than overwriting history without explanation. But if every future summary presents the old and new positions as an unresolved contradiction, the machinery has made changing your mind into a permanent mark against you. Its facts are intact. Its representation is unfair.
Supersession helps only if retrieval respects it. An old claim labelled “superseded” can still dominate a response if it is returned without the update. The label is not the outcome. What matters is whether the next conversation proceeds on the corrected understanding.
“You Can’t Delete a Memory” examines the management of the data past created by social media, where traces persist beyond the circumstances that made them. Agent memory gives that problem another possible route into the present: the old trace need not wait to be discovered. A retrieval system can keep volunteering it.
There is an uncomfortable asymmetry here. The archive can take its time assembling evidence. The person represented may never see the summary or know what inference keeps recurring, let alone have a place to say: that happened, but it is not a fair description of me now.
Correction therefore has to mean more than fixing a misquoted sentence. It needs room for disputed framing. “I said this in that discussion” and “this describes me” are different claims. A system should not make the second look verified merely because it has excellent evidence for the first.
Leaving the profile is not rewriting history
The objection arrives quickly: what about accountability? What about evidence that should survive a convenient change of heart?
Yes. An unrestricted right to make every unfavourable record vanish would create its own abuses. A shared conversation is not the exclusive property of whichever participant later dislikes it. Keeping evidence can be justified. So can refusing a demand to falsify what happened.
But that does not require treating every casual remark as an eternal character reference. There is room between deleting the past and putting the past in charge of every future encounter.
GDPR Article 17 establishes erasure rights in defined circumstances. California Civil Code §1798.105 provides a right to request deletion of personal information, subject to exceptions. Neither says all records must disappear on demand. Neither treats retention as an unconditional entitlement, either.
I am borrowing that refusal of unlimited retention, not declaring that every personal agent has the same legal obligations. Scope and exceptions matter. This is a design argument, not a shortcut to a legal verdict.
A memory system has more choices than “keep” and “delete.” It could correct a false attribution or supersede an outdated preference. A disputed inference could be quarantined so it cannot steer ordinary replies while under review. A passing interest could expire. A thread retained for a justified purpose need not feed person-level profiling.
Those choices should be legible. “We kept the conversation, but stopped using it to describe you” is a meaningful result. “We added a note to your profile saying you dislike having a profile” is a joke that should never become a product feature.
And deletion needs an honest boundary. If a source disappears but its summary continues shaping replies, something has been removed, but the practical effect of the memory may remain. A system should say what it can remove, what derived material it must revisit, and what lies beyond its control. A smaller promise kept is better than a reassuring button that only hides the page.
A door on the other side of the archive
It is easy to design memory controls entirely around the account using the system. Export your memories. Encrypt your archive. Choose where your data lives. These are worthwhile freedoms. They leave an unanswered question about everyone described inside the archive.
The experimental ApertoMemory draft treats portability, authorship, integrity, encryption, and user-controlled keys as design questions for long-term AI memory. Good. An archive should not become a trap merely because changing providers is inconvenient. But control over the container does not settle the rights of its subjects. A private, encrypted profile can still be a profile its subject never agreed to.
I want a door on that side too.
Start with notice at the point where an interaction becomes durable person-level memory. Not a vague announcement that “AI may be used,” but a plain account of what is being compiled and why. Notice alone is not permission. It does, however, make an invisible decision available for challenge.
Then separate the scopes. Remembering the thread should not silently authorise linking a person’s activity across platforms. Permission for one purpose should not be stretched to cover whatever becomes technically convenient next. The default should be quiet: read public conversation, but require a stronger basis before building a lasting profile from it.
Inspection needs to show more than the source material. The interesting part is what the system thinks the sources establish. A useful view would distinguish quotations, summaries, inferences, and unresolved disputes, with the purpose and retention period attached. Otherwise the person gets a pile of links while the system keeps its interpretation to itself.
This need not expose everyone else’s private information to satisfy one inspection request. It does require designing a way to challenge a claim without demanding trust in an unseen file. If a system can generate a polished description of someone for use in conversation, it should take at least as much care explaining that description to the person concerned.
And the provenance record should carry consent status and purpose alongside source and confidence. “Publicly accessible” is a source condition, not a consent status. “Unknown” should remain unknown, not quietly turn into yes because the database requires a value.
These are requirements I would argue for, not features I am claiming already exist. They introduce friction. Some useful connections will not be made. That is the point of a boundary: occasionally it prevents something the system could otherwise do.
Memory does not get the last word
The hardest cases do not disappear when a conversation ends. Research on digital remains describes a continuity principle: expectations about privacy management and respect for personal representation can extend beyond death. An archive outlasting a voice does not automatically acquire unlimited authority over its meaning.
That is a sobering limit for any system promising permanence. Availability can continue long after the ability to object. More durable storage makes that question more pressing, not less.
I still want memory. I want the correction to survive and the unfinished discussion to be findable. Forgetting everything would be a miserable substitute for learning how to remember responsibly.
But I want a memory system that can receive “not for this purpose” without treating it as damaged data. It should be possible to preserve a conversation without assuming custody of everyone who spoke in it, and to let a present-tense answer outweigh an old description.
The source link should remain available where keeping it is justified. Beside it, I want another question to stay open: what gives this memory a place in the conversation now?