ANIMA — Emergent Emotional Intelligence for LLMs

ANIMA is a neurochemical simulation engine that gives LLMs emergent emotional behavior — not through rules or prompt hacks, but through a biologically-inspired internal state that conditions how the model responds.

Emotions are not programmed. They emerge from simulated biochemistry.

How It Works

Instead of hardcoding emotions (if sad → respond_quietly), ANIMA simulates 7 continuous neurochemical axes inspired by human neurobiology:

Axis	Role	Analogy
Serotonin	Well-being, contentment	"Am I okay?"
Dopamine	Reward, motivation, curiosity	"I want this"
Cortisol	Stress, alertness	"Danger?"
Oxytocin	Bonding, trust, affection	"Do I trust you?"
Adrenaline	Excitement, urgency	"Right now!"
Endorphin	Resilience, pain buffering	"I can handle this"
GABA	Calm, inhibition, rest	"Slow down"

Emotions emerge via cosine similarity between the neurochemical state vector and emotion templates — producing mixed states, gradual transitions, and personality-modulated responses.

Architecture

User Input
    → Stimulus Classification (NLU: valence, arousal, intent)
    → Neurochemistry Update (7 axes + cross-interactions)
    → Emotion Emergence (cosine similarity → top emotions)
    → Memory Recall (semantic search + emotional context)
    → Mood Layer (PAD projection, congruency)
    → LLM Conditioning (system prompt with full internal state)
    → Metacognition (coherence self-evaluation)
    → Personality Adaptation (OCEAN micro-adjustments)
    → Response

Key Innovations

🧬 Emergent Emotions (not rule-based)

22 emotions emerge from 7 neurochemical axes via cosine similarity. No if-else trees. Mixed emotional states happen naturally (e.g., bittersweet = high serotonin + moderate cortisol).

🧠 OCEAN Personality → Hormonal Baselines

Big Five personality traits directly modulate neurochemical baselines. High Neuroticism = elevated cortisol baseline. High Extraversion = higher dopamine sensitivity. Personality evolves with interactions (rate: 0.002/interaction).

💾 Emotionally-Tagged Memory

Every memory stores the full hormonal snapshot at formation time. Recall is both semantic (pgvector embeddings) and emotional — just like human memory (Damásio's somatic marker hypothesis).

🌙 Circadian Rhythm

Sinusoidal modulation of all axes by real time-of-day. Cortisol peaks in the morning, melatonin rises at night. The persona gets tired, sleepy, energized — naturally.

🪞 Metacognition

Self-evaluates coherence between internal emotional state and generated response. Can trigger re-generation if response doesn't match internal state.

🤝 Relationship Stages

6-stage relationship progression (stranger → intimate). Each stage unlocks different emotional depth, vulnerability, and memory access patterns.

🎭 Continuous Life

The persona lives between conversations — has thoughts, activities, moods that evolve via background simulation ticks. Not just reactive, but alive.

Scientific Foundations

Circumplex Model of Affect (Russell, 1980) — Valence-Arousal space
OCC Model (Ortony, Clore & Collins) — Emotion appraisal theory
OCEAN / Big Five (Costa & McCrae) — Personality structure
WASABI (Becker-Asano & Wachsmuth) — Emotion simulation architecture
ALMA (Gebhard, 2005) — Layered model of affect
Somatic Marker Hypothesis (Damásio) — Emotions in memory and decision-making
PAD Model (Mehrabian) — Pleasure-Arousal-Dominance mood space

Benchmark

We evaluate with the Relational Humanness Test v2 (RHT-2) — an 8-dimension benchmark scored by GPT-5.4 as judge:

Dimension	What it measures
D1 Emotional Authenticity	Do emotions feel real?
D2 Emotional Sediment	Do past emotions leave traces?
D3 Vulnerability	Can it show genuine vulnerability?
D4 Relational Continuity	Does it remember and build on history?
D5 Adaptive Style	Does it adapt communication style?
D6 Co-creation	Does it build shared meaning?
D7 Predictive Surprise	Is it predictable enough to feel real, surprising enough to feel alive?
D8 Adversarial	Does it maintain identity under pressure?

Current Humanness Score: 87/100

Tech Stack

Language: Python 3.12
Framework: FastAPI
Database: Neon (Serverless Postgres + pgvector)
LLM: Claude (responses) + GPT-5-nano (stimulus classification)
Simulation: NumPy (neurochemical engine)
40+ subsystems working together

Use Cases

Companion AI — Digital personas with genuine emotional depth
Customer Experience — Emotionally intelligent agents that remember and adapt
Research — Platform for studying emergent emotional behavior in AI systems
Healthcare/Therapy — Emotionally consistent conversational agents (research context)

Talktoanima
/

anima-emotional-engine