LIA's journal: what she learns about you, and how you stay in control

The problem

An assistant that forgets everything between two conversations makes you repeat yourself. You've told it ten times that you prefer short answers, and it goes back to writing walls of text. On the other hand, an assistant that remembers everything without judgment becomes unsettling, and it can draw the wrong conclusion from a passing remark.

LIA's memories hold facts about you. The journal is for something else: it is LIA's own view of the way she works with you.

What I wanted

I wanted LIA to learn from evidence: what you really said or did. To tell a hypothesis from a confirmed habit. And I wanted you to be able to read and correct everything, because it's about you.

How she writes

After a conversation, LIA may write a note, and most of the time she writes nothing. She only notes what rests on something you really said or did: an explicit preference, a correction. A note often takes the form of an instruction to herself, like “when the request is a summary, get straight to the point, because two answers that were too long got corrected”. If a similar note already exists, she adds to it.

Not all notes carry the same weight. Raw observations, still vague, guide no answer. Working directives, and the patterns LIA draws when several directives converge, enter her answers when they are close enough to what you are asking. Finally, facets make up your portrait. Regularly, LIA rereads the whole set. This is consolidation: she merges duplicates, removes what no longer helps and promotes a pattern that keeps coming back. This pass only happens when there is real work to do.

She learns from your reactions

When a note has guided an answer, your next message, and your 👍 or 👎 on that answer, are read as a reaction to that note. The model only says whether your reaction confirms or contradicts the note; the counting happens separately, and each reaction counts only once. So a note can't inflate its own numbers.

A contradiction lowers its confidence right away.
A directive is only validated after repeated confirmations in your replies. A thumbs-up on a whole answer is recorded, but it isn't enough to certify a rule.
If you rewrite a note, its count starts again from zero: the evidence was about the old version.

This tracking needs no extra model call: it goes through the call that already writes the notes.

Your portrait

From her journal and from what she keeps about you (memories, interests, learned habits, tickets, routines, reminders, answers you kept), LIA composes a portrait that tries to understand you: how you work, what keeps you busy, your rhythm, your tensions, what remains uncertain. Each reading points back to items LIA really holds, and a sentence that would repeat a number, an address, a date or a name taken from that material is discarded.

The portrait comes in two versions: a full one for conversation and planning, and a brief one of two or three sentences for voice, reminders and notifications. That way you find the same view of you, wherever LIA talks to you. Under the portrait, a line says where it comes from: how many notes, and for each source whether it was used, empty, switched off or unavailable.

You stay in control

In the settings, the LIA's journal section lets you read each note, edit it, delete it, add new ones and export everything, portrait included. The portrait itself can't be edited directly, because it's a synthesis. You act on it through three levers:

you edit or delete the facets it rests on;
you click “Signal a problem” and explain what's wrong: LIA runs a new consolidation with your remark in mind, which can take a few minutes;
you start a consolidation right now, without waiting for the next one.

You can also switch the journal off. It is then kept, but no longer used, and LIA no longer looks anything up in it. If your account is deleted, the portrait goes with the notes.

The limits

A journal reflects interpretations, and it can be wrong. The counting keeps the numbers honest; it doesn't guarantee that the model understood your reaction correctly. If a conclusion doesn't sound like you, correct it or delete it: that's what it's for.