Anima — Ollama Integration

An emotionally aware AI agent with a simulated inner life, running fully locally via Ollama.

---

What is this?

Anima is a Python implementation of an AI agent with a modelled emotional system.
The agent doesn't just generate text — it processes stimuli, forms memories, builds an internal narrative, and only then responds. Every LLM reply is shaped by the agent's current emotional state.

The project is grounded in several theoretical frameworks:

Concept	Implementation
VAD (Valence-Arousal-Dominance)	3D emotional state space
Big Five / OCEAN	Personality model
IIT (Integrated Information Theory)	Approximation of φ-integration
Predictive Processing (Friston)	Prediction error & free energy
Plutchik's Wheel	Eight basic emotions
Confabulation	Probabilistic false memory generation

---

Architecture

AnimaOllama                   ← main agent class
│
├── OllamaBridge              ← REST client for Ollama
├── Personality               ← Big Five traits + confabulation rate
│
├── AdaptiveEmotionMap        ← VAD → emotion map (learns over time)
├── AssociativeMemory         ← associative memory (cosine similarity)
├── IITModule                 ← φ computation (Tononi)
├── PredictiveProcessor       ← prediction error + free energy
├── HomeostaticDrive          ← homeostatic drives & needs
└── ExistentialNarrative      ← first-person inner narrative

Stimulus Processing Loop

stimulus (Dict)
    ↓
[Confabulation?]   ← probabilistic false memories
    ↓
[Memory Resonance] ← past memories amplify current stimulus
    ↓
[Apply + Decay]    ← reactors: tension, arousal, satisfaction, cohesion
    ↓
[VAD Vector]       ← valence = satisfaction - tension, etc.
    ↓
[Emotion ID]       ← nearest emotion in VAD space
    ↓
[Filter]           ← contextual transform (Sadness+tension → Defensive Anger)
    ↓
[IIT φ]            ← entropy of whole vs. parts
    ↓
[Prediction Error] ← surprise level, free energy
    ↓
[Drives]           ← homeostatic needs
    ↓
[Narrative]        ← first-person internal narrative
    ↓
state Dict → OllamaBridge.build_system_prompt() → LLM → response

---

Installation & Setup

1. Install Ollama

# Linux / macOS
curl -fsSL https://ollama.com/install.sh | sh

# Windows — download the installer from https://ollama.com

2. Pull a Model

ollama pull llama3        # recommended (~4.7 GB)
# or
ollama pull mistral       # (~4.1 GB)
ollama pull gemma2        # (~5.4 GB)
ollama pull phi3          # (~2.3 GB, fast)
ollama pull qwen2.5       # (~4.4 GB)

3. Install Python Dependencies

pip install requests numpy

4. Run

python anima_ollama.py              # uses llama3 by default
python anima_ollama.py mistral      # specify a different model
python anima_ollama.py phi3         # lighter model

---

Interactive Commands

Command	Action
(any text)	Regular message — agent responds based on current emotional state
/stress	Increase tension and arousal
/relax	Reduce tension, calm down
/connect	Feel support and connection
/shock	Sudden strong stressor
/state	Display current emotional state
/export	Save state log → state_log.json
/reset	Reset all reactors to baseline
/models	List available Ollama models
/quit	Exit

---

Usage as a Library

from anima_ollama import AnimaOllama, Personality

# Configure personality
persona = Personality(
    neuroticism=0.7,       # prone to anxiety
    extraversion=0.4,      # introverted
    agreeableness=0.8,     # very cooperative
    conscientiousness=0.6,
    openness=0.9,          # highly curious
    confabulation_rate=0.5,
)

agent = AnimaOllama(
    personality=persona,
    ollama_model="mistral",
    temperature=0.8,
)

# Apply a stimulus manually
state = agent.experience({"tension": 0.3, "arousal": 0.2})
print(state["primary"])      # → "Anticipation"
print(state["narrative"])    # → "Something is about to happen..."
print(state["phi"])          # → 0.42

# Get an LLM response shaped by the current state
response = agent.chat(
    "How are you feeling?",
    stimulus={"cohesion": 0.1},
)
print(response)

# Export the session log
agent.export_history("my_session.json")

---

State Dict Structure (experience() → Dict)

{
    "primary":          "Anticipation",          # dominant emotion
    "blend": [
        {"name": "Anticipation", "intensity": 1.0},
        {"name": "Trust",        "intensity": 0.3},
    ],
    "vad": {
        "valence":   0.25,   # -1 (negative) ↔ +1 (positive)
        "arousal":   0.38,   # -1 (calm) ↔ +1 (excited)
        "dominance": 0.42,   # -1 (powerless) ↔ +1 (in control)
    },
    "reactors": {
        "tension":      0.22,
        "arousal":      0.31,
        "satisfaction": 0.48,
        "cohesion":     0.55,
    },
    "phi":              0.38,   # IIT φ-integration
    "phi_label":        "low integration",
    "prediction_error": 0.14,   # surprise level
    "pred_label":       "minor deviation",
    "free_energy":      0.09,
    "surprise_spike":   False,
    "dominant_drive":   None,   # or "cohesion" / "tension" / etc.
    "narrative":        "Something is about to happen. Attention directed forward.",
    "memory_resonance": 2,      # number of activated memories
}

---

Reactors Explained

Four internal reactors define the agent's current state:

Reactor	Baseline	Description
tension	0.2	Internal stress, strain
arousal	0.2	Activation, excitability
satisfaction	0.5	Contentment, calm
cohesion	0.5	Sense of connection, belonging

VAD vector is derived from reactors:

valence   = satisfaction - tension
arousal   = arousal + tension * 0.3
dominance = cohesion + (satisfaction - tension) * 0.5

---

Personality (Big Five)

Parameter	Effect
neuroticism	Amplifies stress response; Sadness → Defensive Anger
extraversion	Amplifies arousal response
agreeableness	Amplifies cohesion; Anger → Assertiveness
conscientiousness	Speeds up emotional decay (return to baseline)
openness	Increases sensitivity to surprise
confabulation_rate	Probability of generating false memories

---

Emotion Filtering Rules

The agent does not always express its raw primary emotion. Context-dependent transforms apply:

Raw Emotion	Condition	Expressed As
Sadness	High tension	Defensive Anger
Fear	Low cohesion	Numbness
Anger	High cohesion + high agreeableness	Assertiveness
Joy	Low cohesion	Relief
Anticipation	High neuroticism	Anxiety
Anger	Low neuroticism	Determination

---

How the System Prompt Works

The agent's emotional state is injected into the LLM system prompt as an <inner_state> XML block. The model is instructed to express the state through language only — never to quote numbers or tag names directly. Guidelines tie specific state values to tonal adjustments:

• High arousal → shorter, energetic sentences
• Negative valence → cautious, withdrawn tone
• High prediction error → express confusion naturally
• Dominant drive cohesion → express need for connection

The prompt is written in English for best cross-model compatibility, but the agent responds in whatever language the user writes in.

---

Requirements

• Python 3.9+
• numpy
• requests
• Ollama running locally

---

License

This project is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0).

English Summary:

• Personal & Academic Use: You are free to use, modify, and build upon this code for non-commercial research and personal projects.
• Commercial & Corporate Use: Strictly prohibited. Any use by for-profit organizations, or integration into commercial products/services, requires a separate commercial license and written permission from the author.
• Attribution: You must give appropriate credit to the original author.

Copyright (c) 2026 [Stell]