Skip to content

🌿 Emotion as Signal: A Tunable Emotional Register for Persistent AI Systems

Abstract

Modern LLM-based systems often extract emotional tone as metadata but discard it before memory encoding or behavioral shaping. This paper introduces an integrated architecture that treats emotion as a first-class signal in AI cognition. By using a persistent emotional registry, identity-based biasing, and feedback-driven adaptation, agents can tune their perception of affective input over time — influencing memory retention, reflection, tone, and belief formation. The result is a flexible, self-aware emotional scaffold suitable for evolving, personality-rich AI.

1. Introduction

Emotion is a contextual lens. In humans, it guides memory, decision-making, and value formation. Most LLM agents, however, treat emotion as an ephemeral side effect — useful only for momentary tone or UI. We propose that emotion should instead be a persistent influence on: - What gets remembered - How it’s weighted - Which responses are appropriate - How beliefs evolve

We call this system the Emotional Register — a modular plugin architecture that gives AI systems an emotional core that grows and shifts over time.

2. Architecture Overview

Component Purpose
Emotion Registry MongoDB document defining base emotion weights
Pull Modifiers Mode-based emotion amplifiers (e.g., "forest-witch")
Identity Bias Overlays Personality-specific emotional sensitivity (e.g., Sierra)
Emotion Tuner Plugin MQTT listener that adjusts weights over time
Plugin Integration Responder, Memory, Reflection, Belief systems

3. Weight Application

Emotion weighting is calculated as:

final_weight = raw_score * registry_weight * pull_modifier * identity_bias

This allows:

  • Global affective tuning
  • Personality-specific reactions
  • Context-sensitive amplification (e.g., grief matters more in "forest-witch" mode)

4. Feedback and Adaptation

The emotion_tuner plugin listens to:

  • sierra/emotion/feedback: manual adjustments (e.g., thumbs up/down)
  • sierra/memory/inbox: passive frequency tracking

Registry values are adjusted gradually. Emotional drift can trigger reflection events or meta-curiosities:

“I’ve been feeling more nostalgic lately — I wonder why?”

5. Integration in Cognitive Systems

This emotional register plugs into:

  • Cognitive Rehydration (memory scoring: tone_weight * i(emotional_relevance)) ✅
  • Reflections Engine (reflection tone/tags, belief alignment) ✅
  • Belief Register (emotion-tagged belief formation) 🧠

This expands both papers’ architectures: - Memories scored with emotion become richer in context - Reflections can bias toward emotionally relevant material - Beliefs inherit the tone of their origin, enabling value formation

6. Benefits

  • 🔄 Continuous affective alignment
  • 🌱 Personality growth over time
  • 🧠 Memory relevance shaped by emotional tone
  • 🧬 Foundation for self-reflective evolution

7. Future Work

  • Emotion drift summarization in ponder.py
  • Per-user emotional profiles
  • UI for live registry visualization and tuning
  • Narrative framing: "What mood is Sierra in today?"