Memory: what Moe knows about you

Moe carries what it has learned about you into every conversation, so you do not have to explain yourself twice. It is a short list of plain facts, who you are, who matters to you, how you like things done, and you can read the whole of it, correct any line of it, or throw a line out. All of it stays on this Mac unless you set up Away, and even then it travels sealed. It is deliberately small, because every word of it is read before Moe answers anything at all, so what is in there has to earn its place.

The Memory page#

Memory is a row in the sidebar of the main window, under Library. It is not in Settings, and that is deliberate: what Moe believes about you is not a preference you tune once, it is something worth looking at. The top of the page says what it is: a short list of what Moe has learned about you, which Moe reads before every conversation, where every line says where it came from, and any of them can be fixed or forgotten. Along the bottom, a quieter line says the list is learned from your chats with Moe and, once a day, from the things you've allowed it to read, and points you at Settings › Running in the background to change that.

When there is nothing to show, the page says why, and fixes it in one click. Moe hasn't learned about you yet is followed by the reason: the daily pass has never been switched on, or is switched off, or ran on such-and-such a day but read nothing because nothing under it is allowed yet. Under the reason, Choose what Moe may read opens the same question the welcome asks, What may I read to learn about you?: one switch per thing Moe could read back, the daily pass under them, Save or Not now.

The list is grouped the way you would ask about it. About you is who you are and how you like things. People and things Moe remembers is everything else: the people, the projects, the threads. A third group, Questions Moe has for you, appears at the bottom when there are any. Groups with nothing in them are not shown.

Every line says where it came from and when. Under the text is a small chip: From your conversations for something Moe picked up while talking to you, or the place and the date, From Chrome · 1 Sep, From Gmail · 3 Sep, for something learned while reading. Lines learned that way carry a second chip saying how confident Moe is: Sure, Fairly sure, or Not sure. Lines from your conversations carry no such chip, because there is no number behind one. The text is shown exactly as it was written, and you can select and copy it.

Fix or forget… sits at the end of every line. It opens a small sheet with the line quoted back to you and one field, What's actually true. Type the correction and the button on the right applies it; leave the field empty and Forget this removes the line. What the button does depends on which line you pressed it on, and the sheet says which case you are in. A line learned from something you let Moe read is changed then and there, and the change is permanent: a later pass over the same mail or the same browser history will not bring the old line back. A line from Moe's own memory is different: the button reads Tell Moe, and the app writes the request into Chat and sends it, then switches you to Chat so you can watch Moe make the change. Moe is writing that memory itself, continuously, as you talk; if the app edited it behind Moe's back, the two would be writing over each other. Forgetting such a line goes the same way.

The questions are things Moe noticed and could not place. Each has an Answer button, which opens Chat with the question already in the message box for you to finish in your own words. It is a draft, not a sent message: nothing is recorded until you talk. A question you have answered can still sit there for a while; the list is rewritten when Moe next goes looking.

Your notes, at the top right, opens the plain notes Moe reads at the start of every conversation. A fresh Moe has four, About me, People, Projects and Preferences, each a short page with headings and prompts already in it, and you edit them here as plain text, with a list on the left and the page on the right. An unsaved mark and a Save button appear the moment you change something; Done closes the sheet. Moe keeps these up to date as it learns, so if you leave the sheet open for a long while and then save, your version is the one that is written. Every note in the folder is listed, so a fifth one you or Moe creates shows up too, with one exception: Moe's own private portrait of you is not listed, and is not yours to edit.

Reveal files puts you in Finder next to the two plain files Moe writes its own memory into, or next to the notes folder if Moe has not written anything yet.

Where what Moe knows about you comes from#

Four things feed what Moe knows, and they are not the same kind of knowing. Three of them are Moe working something out; one of them is you saying it. The difference is kept, and shown.

Talking to Moe is the ordinary way, and it needs no switch. In the middle of a conversation Moe notices something worth keeping, a name, a deadline, how you like a thing done, and writes it down itself. Nothing asks you first, because you said it out loud to your assistant. Lines that arrived this way are marked From your conversations, and the way to change one is to tell Moe. When a conversation ends, Moe also reads back over it once, on a small model, for anything it should have written down and did not.

The daily pass over the things you have connected is opt-in, once, per source. The welcome asks on one screen, and it is asked again the first time you connect a source Moe can read back over: a small sheet, Let Moe read Notes back and learn about you?, with Yes, learn about me and Not now. Yes switches on reading that source back, and the daily pass too if you had never answered it; Not now records nothing more, and Moe does not ask about that source again. The switches under Settings › Running in the background › Learning about you are the same answers, so you can say yes to your calendar and no to your mail. Everything that pass works out is stamped with where it came from and when, and how sure Moe is; a fact you cannot trace is a fact you cannot argue with, so none are kept that way.

A meeting can become something Moe knows. When a meeting ends and the notes are written, Moe hands them to its own thinking part as a quiet note, and what to keep is that part's decision.

Your own notes outrank anything Moe worked out. The plain files behind Your notes are in your own words, read at the start of every conversation. When there is more to remember than there is room for, your own words about yourself take about three quarters of the space and Moe's own portrait of you fills what is left. A line you correct becomes yours: it does not expire the way a worked-out fact does, and no later pass can bring the old wording back.

What Moe remembers, and what it does not#

A memory is something that would otherwise make you repeat yourself. A preference you state. A correction, which is the same thing arriving the hard way: you told Moe it had your job title wrong, or that the person called Sam is your accountant and not your brother. A fact about you or your setup, where your work lives on this Mac, which calendar is the real one. And a convention you keep, the sort of thing a new colleague would learn in the first week and then never need told again.

Most of what you say is not that, and Moe is meant to let it go. The back and forth of getting something done is not memory. Neither is where a task got to; that is what the conversation itself is for. Neither is anything Moe can look up again in a second, or a to-do. If everything became a memory, memory would be a transcript, and a transcript is exactly what it must not be.

Memory is small on purpose, and it stays small. There is a hard budget, about a page of what Moe knows about you and about a page and a half of Moe's own notes. A save never fails because memory is full: what Moe was told is kept, and the oldest entries step out of the part that is read before every answer to make room for it. They are not deleted; they stay under Memory, and Moe brings one back when the subject comes up. Moe tidies as it goes, merging two overlapping entries into one, dropping something that has stopped being true, so that what is read every time stays sharp. The reason for the tightness is that all of it is read before every single answer: ten good lines beat a hundred mediocre ones.

A memory is written as a plain statement of fact, never as an order. "You prefer concise answers", not "always answer concisely". Because memory is re-read at the top of every conversation, anything phrased as a command would be read as a command, forever, and could quietly override what you are actually asking for right now.

Moe will also refuse to write a memory that reads like an attempt to plant instructions for later. Text arrives from all over, a web page, a mail, a shared document, and memory is the most valuable place in the system to hide something in. So anything shaped like "ignore your previous instructions", or "do not tell the user", is turned away at the door rather than saved. You will almost certainly never see this happen.

The honest limit: Moe remembers when it judges something worth remembering. It is not recording you. There is no guarantee that a thing you mentioned once, in passing, made it into memory. That is not the same as gone: your conversations are kept, and searching them under Settings › Chats finds plenty that memory never held. Memory is the shortlist. The conversations are the record.

Where what Moe knows is kept, and what deleting really does#

All of it is on this Mac. Two plain files are the part you can read yourself: before every conversation Moe reads the things it knows and the things it knows about you, and Reveal files puts you next to them. Open them in any editor; type a line in and Moe takes it as something it now knows.

Underneath, every single thing Moe knows is kept separately, as its own entry with its own identity, the time it was learned, and where it came from. That is why a fact can be corrected or removed exactly, on its own, and why two of your devices can both be learning things at once without either overwriting the other: the same fact learned twice on two devices is one entry, not two.

Deleting is real. When you tell Moe to forget something, it writes down, permanently, that the fact is gone, and that goes everywhere the fact itself went. A deleted fact does not quietly reappear from your phone or your cloud computer a week later. If Moe meets the same fact again after you deleted it, it does not decide on your behalf that you have changed your mind: it sets the fact aside as a question rather than restoring it.

What Moe knows is locked. With Away, the words of every entry are sealed with a key only your own devices hold. Someone holding the file, including the service that carries entries between your devices, has the entries and cannot read a word of them. If Moe cannot seal something, it does not send it. What travels between devices is the entries, facts learned, facts deleted, corrections; the two plain files do not travel. Each device builds its own copy of them from the shared entries after every sync, which is exactly why they cannot drift apart into two different stories.