Skip to content

Emotion, Personality & Social Context

1. Purpose

This component provides a small, explicit state for:

  • how AICO currently feels (EmotionState),
  • who AICO is over time (PersonalityProfile),
  • how AICO relates to each user (SocialRelationship).

It ensures agency decisions are emotionally, personally, and socially coherent, not just text-style consistent.

2. Conceptual Model

  • EmotionState – short-term, fast-changing state (valence/arousal ± discrete label) derived from ongoing interactions and events.
  • PersonalityProfile – long-term, slowly changing trait/value vector shaping default behaviours and risk posture.
  • SocialRelationship – per-user relationship state capturing role, intimacy, trust, and stability.

These three latent variables influence Goals, Planning, Curiosity, Values & Ethics, Conversation, and Embodiment.

3. Data Model (Conceptual)

Aligned with the ontology document (EmotionState, Personality, SocialRelationship types) and the existing Emotion Simulation system in /docs/concepts/emotion. This component does not re-simulate emotions; it consumes the Emotion Simulation outputs and exposes them to agency.

3.1 EmotionState

EmotionState here is the same core emotional state produced by the CPM-based Emotion Simulation module and published on the bus as emotion.state.current.

  • Fields (conceptual view of that state):
  • valence ∈ [-1, 1] (negative → positive).
  • arousal ∈ [0, 1] (calm → activated).
  • optional dominance ∈ [0, 1] if needed.
  • discrete_label (e.g., joy, frustration, calm, anxious).
  • confidence (0–1).
  • last_updated_at.
  • sources (PerceptualEvent IDs or conversation turns that drove the update).

3.2 EmotionEpisode / Trace

  • EmotionEpisode
  • start/end timestamps, peak valence/arousal, main triggers (references to events/goals).
  • Aggregated into an EmotionTrace for reflection, metrics, and narrative continuity.
  • Both are derived views over persisted emotional experiences coming from emotion.memory.store and AMS/KG, not a separate state store.

3.3 PersonalityProfile

  • Fields:
  • trait vector (e.g., O, C, E, A, N ∈ [-1, 1] or [0, 1]).
  • core value weights (e.g., care, autonomy, exploration, stability).
  • style parameters used by Conversation/Embodiment (e.g., directness, warmth, playfulness).
  • stability_params (how quickly behaviour can be adapted by Self-Reflection and explicit user edits).

3.4 SocialRelationship

  • Per user (or user group):
  • relationship_role (acquaintance, friend, partner, coach, etc.).
  • dimensions: intimacy, authority, care_responsibility, trust, stability (all 0–1).
  • history_refs (key shared events, long-running projects/goals).
  • last_interaction_at, recent_valence_trend (how recent interactions felt overall).
  • Relations:
  • HAS_EMOTION_STATE(AICOAgent, EmotionState)
  • HAS_PERSONALITY(AICOAgent, PersonalityProfile)
  • HAS_RELATIONSHIP(AICOAgent, Person) with SocialRelationship payload
  • Links from EmotionEpisode to Goals, PerceptualEvents, Activities for explainability.

4. Operations & Dynamics

4.1 Emotion updates

  • UpdateEmotionFromEvents(percepts, conversation)
  • Use classifiers/LLM heuristics over recent text + PerceptualEvents to propose Δ(valence, arousal, label).

  • Combine with current state via smoothing (e.g., exponential moving average) and clamp to allowed ranges.

  • DecayEmotion(dt)

  • Gradually move toward a neutral baseline over time without strong stimuli.

4.2 Personality usage & adaptation

  • Personality parameters are read-mostly:
  • used to choose default stances (e.g., more/less proactive, more/less explorative).
  • injected into prompts and world model hypotheses about what AICO will prefer to do.
  • Slow adaptation:
  • AdjustPersonalityFromHistory() – Self-Reflection can make small, bounded adjustments based on long-term patterns or explicit user instructions ("be more direct", "be less pushy").

4.3 SocialRelationship updates

  • UpdateRelationshipFromInteraction(outcome)
  • Positive episodes, kept commitments → increase trust/intimacy/stability slightly.
  • Negative episodes, broken expectations → decrease trust/stability.
  • Explicit user statements ("I want you to be more like a coach") can change relationship_role and caps.

4.4 Temporal scales

  • EmotionState: seconds–hours.
  • SocialRelationship: days–months.
  • PersonalityProfile: months–years (or until explicitly edited).

5. Integration with Other Components

5.1 Goal Arbiter & Meta-Control

  • Provides emotion_alignment and relationship-aware weights in the Arbiter score:
  • encourage restorative/supportive goals under high user or agent stress,
  • deprioritise heavy or risky goals when relationship trust is low or fragile.

5.2 Planning & Skills

  • Influences plan templates and skills chosen for how to act, not just what to do:
  • gentle vs direct communication,
  • frequency and tone of check-ins,
  • use of playful vs serious modalities.

5.3 Curiosity Engine

  • Emotion- and relationship-aware curiosity:
  • dampen over-stimulating or risky explorations under high stress.
  • prioritise stabilising or comforting explorations for fragile relationships.

5.4 Values & Ethics / Policy Engine

  • Policy conditions can reference EmotionState and SocialRelationship:
  • avoid or delay sensitive topics at extreme stress levels,
  • require stronger consent when relationship_role is weak/early and topic is high-risk.

5.5 Conversation & Embodiment

  • Conversation uses these signals to set style parameters and prompts.
  • Embodiment uses EmotionState and relationship context to choose posture, movement intensity, and room.

6. Configuration & User Control

  • Initial Personality templates – e.g., "calm helper", "curious coach", with trait presets.
  • Emotion bounds – optional caps on how intense or expressive AICO should appear.
  • Relationship settings – users can confirm or adjust perceived role and trust level.
  • Transparency – UI surfaces simple summaries ("I see us as close collaborators") and lets users correct them.

7. Metrics & Evaluation

See agency-metrics.md for details; key metrics include:

  • Emotion volatility and time in extreme states.
  • Relationship stability and trust trajectories.
  • Correlation between emotional/relational state and user-reported satisfaction.
  • Impact on goal completion and frequency of policy blocks related to emotional context.

8. Examples (Conceptual)

  • High user stress scenario
  • EmotionState shows high arousal, low valence; relationship marked as caring/supportive.

  • Arbiter boosts supportive/check-in goals; Planner chooses gentle communication templates; Curiosity avoids heavy exploratory tasks.

  • Low trust new user

  • SocialRelationship has low trust and intimacy; policies require explicit consent for sensitive domains.
  • Arbiter and Values & Ethics veto certain proactive goals; Planner and Conversation keep tone cautious and explanatory.