Our multi-agent debates always had good bones. A CFO persona, a Growth persona, a Risk persona — each with a mandate, each arguing a real decision until a recommendation fell out. But they were amnesiac strangers. Every debate, three brilliant advisors met for the first time, disagreed politely, and went home. Nobody held a position long enough to be accountable for it. Nobody remembered being wrong.
This update changes that. PersonAI personas now have an inner life: a mood, emotions, relationships they carry from one debate to the next, and a memory that is shaped by how they feel. It is the largest change to the debate engine since we first shipped it — and it made the boardroom sharper, more honest, and occasionally petty in exactly the way real rooms are.
From stateless critics to characters with a past
Before, each persona was a fresh function call. Now each one has persistent affective state that survives the conversation: a current mood, a temperament it returns to, and a graph of relationships. Every emotional beat — a concession, a clash, a vindication — is written to an append-only event log, so a persona's disposition is something that accumulates rather than resets. Two advisors who have clashed for three debates walk into the fourth already wary of each other.
Feelings, with coordinates
Mood isn't a label we slap on — it's a point in space. We model it with three axes: valence (is this good or bad), arousal (calm or activated), and dominance (in control or on the back foot). A persona drifts through that space as an argument unfolds, then decays back toward its baseline between sessions, at a rate set by how volatile its temperament is.
On top of those coordinates sits a lexicon of roughly thirty-five discrete emotions — from conviction and curiosity to contempt and unease. As a persona reasons, it emits inline signals the engine parses and applies, targeted at whatever it's reacting to. Arousal even feeds back into generation: an agitated advisor is sampled at a slightly higher temperature, so it literally argues with more heat.
A persona can love an idea and distrust a company
Relationships are directed edges, and their target can be almost anything. A persona can feel something about another persona, about you, or about any entity in the debate — a market, a competitor, a pricing strategy. It can be loyal to an idea and cold toward the company championing it. Those bonds are typed, persistent, and — crucially — they change behavior.
We turn affinity into explicit instructions the persona has to reason around. Distrust is told to demand evidence before it agrees. Affection is warned that it's prone to motivated reasoning. A persona doesn't just have a relationship; it has to account for one.
A board that never remembers being wrong is just an expensive echo. The moment an advisor can hold a grudge, it can also hold a conviction — and be answerable for it.
Memory that feels
Recall is now affect-ranked. When a persona pulls context, we score candidate memories on a blend of semantic similarity, recency, salience, and mood congruence — so an advisor in a sour mood surfaces the sharper memories, the way people do. The strongest emotional moments are saved as high-salience memories that resurface debates later, which is how a persona that got burned on an optimistic call once will bring it up the next time optimism is on the table.
The backchannel
Between the opening statements and the first rebuttal, something new happens off the main stage. Personas with a strong bond quietly send each other a single private note — allies coordinate their line of attack, rivals confront each other before the room does. Only they can see it; you see that it happened. It emerges on its own: with no relationships in play it's a no-op, but once bonds form, the debate grows a backstage.
Track records, humility, and self-knowledge
Feelings would be theatre if they had no consequences, so we closed the loop. When you resolve a decision later and tell the system how it actually turned out, every persona that made a prediction is scored. Get it right and it's vindicated — more confident, standing taller. Get it wrong and it's humbled. An advisor that has missed four calls out of five now hedges, and is told, in as many words, to hold its views loosely. One that has been proven right is reminded not to get overconfident.
Each persona also carries a self-model: a short, honest read of its own disposition and blind spots. When it's about to defend something it's attached to, it's instructed to name the bias and reason past it — and the synthesizer now credits the advisors who own their bias while flagging the ones who don't.
Watch it think
None of this is a black box. An interactive constellation renders the whole team as a living graph you can drag, zoom, and switch between three lenses — what everyone feels, what the workspace knows, and what each persona remembers — with a glow where feeling and knowledge overlap. A dedicated Mind view gives every persona a FEELS / REMEMBERS / KNOWS profile and a time scrubber that replays its mood and relationships from the event log. Personas even write short autobiographical chapters as their temperament drifts, so an advisor's arc becomes something you can read.
Why we built it
Not for novelty. A board of advisors is only worth having if it brings continuity, conviction, and the humility to update — the things that make human counsel valuable and that stateless models throw away between every call. Giving our personas an inner life didn't make them nicer. It made them accountable. And an advisor you can hold accountable is the only kind worth listening to.
