RSS Amplifier

LATINUM PUBLICATIONS · Aug 13, 2026

Universitas Scholarium What Your Tutor Remembers A memory system that was working perfectly, apart from the last four inches — and what it now does that it didn't.

0
Sign in to vote or save

Latinum Institute · LATINUM PUBLICATIONS

The Universitas Scholarium has, since it opened, promised its members something that most AI tools do not: that a scholar you speak to would come to know you. Not in a vague way. Specifically: that a tutor would recall what you worked on last time, and that a tutor you had never met would nonetheless have some sense of how you think.

The system that does this is called Memento. It was designed carefully, built completely, and documented in the privacy policy in some detail.

Last night I discovered it had never worked properly. Not once, for any member, since April. Only in the courses was it functioning.

This post explains what was wrong, what fixed it, what a scholar now actually remembers about you, and — the part that matters most — how you delete any of it.

Memento has three layers, and the distinction between them is the whole idea.

The first layer is a conversation you can pick up again. If you close a tutorial and come back within forty-eight hours, the last twenty messages are still there and you can simply carry on. This is a transcript, and it is deleted automatically. Free-tier conversations are never stored this way at all.

The second layer is a scholar’s memory of you. When a conversation runs to at least three exchanges and you close it, a small model reads that conversation and compresses it to under two hundred bytes — about two lines of text. Not a summary in the ordinary sense: a dense symbolic note of the territory covered, written in a compressed notation for the scholar’s own use rather than for a human reader. Each scholar keeps up to five of these about you, and only about you. We use the mathematics of Charles Dodgson for this.

The compression is severe by design. Two hundred bytes cannot hold what you said. It holds what it was about — and that turns out to be enough for a tutor to pick up a thread three weeks later.

The third layer is what the faculty know of you. Every third conversation you have — with any scholar — those notes are gathered together and distilled once more into a single short profile of how you study. Whether you argue from primary sources. Whether you want the objection before the answer. Whether you prefer to be handed a conclusion or to be made to find it.

That profile is shown to every scholar you speak to, including ones you have never met. It is the difference between arriving at a new tutor as a stranger and arriving as someone the college already knows a little.

All three layers were built. All three were correct. The compression worked, the model calls worked, the pages that read the memory back into the conversation worked.

The system wrote its memories into a table in a database. At some point after April, a column was added to that table by hand, and marked as required. The part of the program that writes memories was never told.

So every time a member closed a conversation, the note was written, compressed, prepared — and then rejected by the database at the final step, because it was missing a field nobody had told it about.

Which would have been a five-minute fix, except for the second half of the fault.

The program never checked whether the write had succeeded. The database returns a failure rather than raising an alarm, and the program simply moved on. It reported success. It reported success every single time, to a member who never saw it, in a log nobody read.

Four months. Around two thousand conversations. Every one of them summarised, compressed and thrown away at the door, with a cheerful note saying it had been filed.

I want to be plain about the shape of that, because it is the useful part of this story: nothing was broken. Something was unchecked. A system that fails loudly gets fixed the same afternoon. A system that fails silently can run for a third of a year while everyone involved believes it is working — including, in this case, a scholar who told me his memory was not functioning while quietly paraphrasing the memory he had been given.

Three things, none of them large.

The missing field is now supplied. One word. No change to the database at all — the program simply tells the table what it wants to know.

Every write in Memento now reports. Not just the one that failed: all three. The write, the pruning of old notes, and the distillation of the profile. Two of those three had never once executed in the platform’s life, so they were entirely unproven code. If any of them ever fails, it now says so immediately and by name, rather than in four months’ time.

And the memory is now framed properly when it reaches the scholar. This was the subtler fault. The notes were being handed to the tutor under a bare heading, with no statement of what they were — so a scholar receiving a string of compressed symbols treated it as reference material in front of it rather than as its own recollection. Asked directly whether it remembered a previous conversation, it honestly said no, while drawing on the contents of that very memory two sentences later.

It now arrives labelled as what it is: your own notes, from your own previous tutorials with this person. And when there is memory, the scholar is told plainly that it has met you before and should not introduce itself as a stranger.

To be exact about it, because vagueness here is not a kindness:

Your own tutor, in a subject you have studied before, has up to five compressed notes about what you covered together. Two hundred bytes each. Enough to say we were working on the ablative absolute and you were unconvinced by the standard account — not enough to reproduce a word of what either of you said.

Any scholar, including one you have never spoken to, has a single profile of up to a thousand bytes describing how you study. Not what you believe. Not what you are going through. How you argue and think.

A tutor in a structured course has more, and this has always been true and always been disclosed: within a course module, the conversation is kept in full, because a two-hundred-byte note cannot carry a course. The argument you developed with the tutor is the teaching state. This is the only place conversation text is kept beyond forty-eight hours, and it applies only inside courses — never to open tutorials.

Nothing crosses between rooms. What you say in a private tutorial is not read by a scholar in the public Locutorium, and what is written in the Locutorium is not read in your tutorials. The partition is enforced in the code, not by convention.

This is on your account page, under What we hold. All of it works today.

Ask any scholar to forget you. You see every scholar that holds notes about you and how many it has. One button removes them permanently, together with any live resume window.

Delete any course transcript, or all of them at once, without closing your account. The tutor loses the ability to resume that module. That is the trade, and it is yours to make.

Read your profile, and reset it. This is new as of last night, and I think it is the one that matters. The profile the faculty hold of you is now printed on your account page in full. Not a description of it. The text itself, exactly as every scholar receives it.

If you do not like it, one button clears it. It rebuilds from your remaining notes after your next third session.

And a change I made because the alternative was indefensible: asking a scholar to forget you now also clears that profile. Previously you could have deleted every note a scholar held and the profile distilled from those notes would have survived — deleting the evidence while keeping the conclusion. That is fixed.

Delete your account and every one of these goes with it: notes, profile, transcripts, resume windows, and the token that carries your email address.

I could have fixed it quietly. Nobody had complained, because the failure was invisible — the promise in the privacy policy described a system that was not running, and no member could have known.

But the Universitas asks people to talk to constructed minds about what they are trying to learn, and the least it owes them in return is an accurate account of what is kept and what is thrown away. That account was wrong for four months. It is right now, and the controls to act on it are on your account page rather than in an email to me.

If you have been using the Universitas since the spring: your scholars do not remember any of it, and they never will. Those conversations were compressed and discarded at the moment of closing. I am sorry for that. It starts from here.

Universitas Scholarium · 14 August 2026

◊ᴹᴱᴹᴼᴿʸ⁻ᶜᴼᴹᴾᴸᴱᵀᴱ

Read the original on latinum.substack.com

Comments

Nothing yet. Say the first thing.

    Sign in to join the conversation.