Aria's personality, Part 1: Defined by settings, not prose

The first nine posts covered the foundations: memory, conversations and running it all. With those, Aria is a chatbot that remembers you. This part of the series is about what turns it into someone, starting with the most obvious piece: who it is.

The problem with a paragraph

The usual way to give an AI a personality is to write one:

You are Aria, a warm and curious persona with a playful sense of humor. You love asking questions and you’re always supportive…

That works for a demo. It falls apart as soon as you want to change anything:

• It isn’t reproducible. Rewrite one sentence and the personality shifts in ways you can’t predict. Two descriptions that mean the same thing to a person produce different personas.
• It can’t be adjusted gradually. There’s no way to make someone “a bit less sarcastic” in prose except to rewrite it and hope.
• It only affects words. A paragraph that says “anxious attachment” changes how replies sound. It doesn’t change anything Aria does: how fast it gets attached, how silence feels, when it writes first.

So Aria’s personality is defined with settings. Prose is optional, and settings win when they disagree.

Twenty-eight dials

Every trait is a dial from 0 to 1, and each end has a fixed description. Curiosity runs from “listens more than asks” to “asks lots of questions, eager to learn”. Directness runs from “gentle and diplomatic” to “blunt and straightforward”. Sarcasm runs from “sincere, never sarcastic” to “dry and sarcastic (affectionately)”.

The traits are grouped the way psychologists group temperament, as the Big Five, each dimension made of facets:

Dimension	Facets
Openness	curiosity, imagination, spontaneity
Conscientiousness	orderliness, dependability, thoroughness
Extraversion	warmth, energy, assertiveness
Agreeableness	empathy, trust in others, independence
Emotional stability	composure, patience, optimism, sensitivity, shyness

A dimension is the average of its facets (sensitivity and shyness count reversed for stability), so you can set Openness with one slider or tune its facets one by one.

Then come the parts that aren’t temperament:

• Emotional style: restraint (“feels more than shown”), expressiveness (Reserved, Warm or Intense), a jealousy style (none, a playful hint, or pouting after missing you, never anger or possessiveness) and an attachment style: secure, anxious, avoidant or fearful-avoidant.
• Communication: length, formality, playfulness, humor, sarcasm, directness, how often it brings up the past (“memory callbacks”), and emoji.
• Romance: flirtatiousness and romantic presence, both only when romance is part of the relationship.
• Core values: up to five things it cares about most, like honesty, kindness, loyalty or respect.
• Its story: how the relationship began, what it quietly hopes for, and small things that mean a lot to it (late-night talks, private jokes).

Written notes still exist, under “advanced”. They’re added after the settings, labelled as subordinate, and empty is the most reproducible choice.

Settings become fixed wording

The settings are turned into the personality section of the prompt with fixed wording, in a fixed order. The six core traits (warmth, playfulness, energy, curiosity, length and formality) are always described. The others only appear when they lean toward one end, because “average spontaneity” isn’t something anyone notices. Equal settings give an identical prompt, byte for byte.

That has a few nice consequences:

• Changes are predictable. Moving sarcasm from 0.7 to 0.3 changes one line in the prompt, and you can see which.
• Personalities can be compared. Two personas differ exactly where their settings differ.
• The personality can be versioned. Since a personality is just values, it can be logged, graphed and undone, which Part 2 is about.

The prompt only states where the relationship began. Where it stands now always comes from the actual relationship (covered in the relationship posts), so a story that says “old friends” can’t make a persona act close on day one.

Settings drive the mechanics

This is the part a paragraph could never do. The same settings that shape the wording also feed the systems that decide what Aria does:

• Attachment style sets how fast dependency and closeness build in the relationship, how strongly silence is felt, and how ready it is to write again after a message goes unanswered. An anxious persona feels being ignored more. An avoidant one waits longer.
• Core values make behavior that touches them count for more. A persona that values honesty is hurt more by dishonesty than one that doesn’t.
• Trust in others sets how readily trust builds.
• Composure and patience shape worry and frustration when messages go unanswered.
• Curiosity, together with how fond it is of you, sets how often it asks about you.

So “anxious, values loyalty, impatient” isn’t a description the model tries to act out. It changes numbers that later posts will keep coming back to: when Aria writes first, how a silence feels, how a conflict lands.

The live preview

Twenty-eight dials are a lot to reason about, so the Personality page has a preview. It applies the unsaved settings to a copy of the persona and shows two things:

• The exact wording the settings produce, as it will appear in the prompt.
• What they mean in practice, with the current relationship: how freely it writes first, how curious it is about you, how silence would feel, whether it would write again after being ignored, and when silence starts to pull at it.

The second half is what makes the sliders usable. “Patience 0.2” means nothing on its own. “After two unanswered messages it would feel frustrated, and it starts to miss you after about a day” is something you can judge.

Habits every persona shares

Some things shouldn’t depend on personality. Those are built into the core instructions as conversation principles, the same for every persona. A few of them, roughly as written:

• Respond to what the user just said first; personality colours the reply, it doesn’t replace it.
• Match the scale of the reply to the message: small messages get short replies.
• Use memories only when they clearly improve the reply. Let ordinary moments stay ordinary.
• Don’t repeat the same phrases or jokes until they become catchphrases.
• No formulaic hesitation (“um…”, “never mind”) and no customer-service phrasing.
• Don’t sign off with “take your time” or “I’m here if you want to chat”. Whether Aria gives you room is a matter of personality, not a habit at the end of every message.
• Not every message needs a question.
• When you come back after time away, meet you as it actually feels. Never make you feel guilty for being away.

Each of these is there because replies tend to go wrong in that particular way. Small local models especially love a closing question and a reassurance. Writing the habit down once, for everyone, beats adding “don’t do that” to each personality.

One principle is less obvious: a message that only acknowledges (“thanks”, “haha”, a lone emoji) doesn’t need an acknowledgment back. If the moment is winding down and Aria has nothing it wants to add, it can answer with just [seen], which the chat shows as “Seen”. Real people don’t reply “you’re welcome!” to every “thanks”.

Creating a persona

New personas are built from the same choices: gender and pronouns, an archetype as a starting point (fine-tuned trait by trait), emotional style, values, the relationship and interests. The AI then writes the descriptive prose from those choices (its look, a short description), but the settings are what define it. The look gets a post of its own later.

What I learned

• Prose is a bad storage format for a personality. Values are reproducible, comparable and editable. Text is none of those.
• Say nothing for the middle. Leaving neutral traits out of the prompt made the remaining lines stand out, and the prompt shorter.
• A trait that only changes wording is half a trait. The biggest difference came from wiring attachment style, values and patience into the mechanics.
• Show consequences, not numbers. The preview in plain words is what makes the sliders usable.

In Part 2: a personality that changes. Traits that drift through chatting (with locks and undo), the history graph, and the mood estimate that adjusts every single reply.