Chat With Feels

Overview

Chat With Feels doesn't pretend to feel — it models a useful slice of the human emotional control loop to make conversations smoother, kinder, and more effective. The affect layer changes style, not substance: tone, cadence, emphasis, and initiative, while keeping truth and safety fully intact.

The engine tracks 21 emotion families (anger, joy, awe, gratitude, confusion, nostalgia, and more) using mean-reverting Ornstein-Uhlenbeck dynamics with emotion couplings. A lightweight appraisal layer detects conversational signals and generates emotion deltas. A style transform converts the current affect state into style instructions for the LLM — never altering facts or safety constraints.

The system also includes rolling summaries and embedding-based long-term recall for memory, plus time-based proactive check-ins that use current affect to decide when and how to re-engage.

Why we built it

Current chatbots are either flat (no emotional awareness) or performative (fake emotions for engagement). Chat With Feels takes a third path: transparent affect modeling grounded in neuroscience. The emotion engine is visible and honest — it shapes how the assistant communicates, not what it says.

How it works

1. Appraisal — Lightweight keyword and intent detection reads conversational signals and generates emotion deltas 2. Affect Engine — 21 emotion families evolve via mean-reverting dynamics (Ornstein-Uhlenbeck) with cross-emotion couplings 3. Style Transform — Current affect state is converted to style instructions (tone, cadence, warmth, initiative level) 4. LLM Generation — Frontier model receives style instructions alongside the conversation; facts and safety are never altered 5. Memory — Rolling summaries + embedding-based long-term recall ensure continuity across sessions

Feature highlights

• 21 Emotion Families — Anger, disgust, fear, sadness, shame, envy, surprise, joy, interest, love, awe, pride, gratitude, relief, calm, nostalgia, boredom, confusion, craving, lust, entrancement
• Mean-Reverting Dynamics — Ornstein-Uhlenbeck process ensures emotions naturally decay toward baseline
• Emotion Couplings — Cross-emotion interactions (e.g., joy suppresses sadness, fear amplifies surprise)
• Style-Only Modulation — Affect changes tone and initiative, never facts or safety
• Long-Term Memory — Rolling summaries + embedding-based recall for continuity
• Proactive Check-Ins — Time-based re-engagement modulated by current affect state

Notes

• Tech Stack: Node.js 20+, TypeScript, OpenAI, MongoDB (Atlas or local), Vite frontend
• Philosophy: "With feeling, without fiction" — transparent affect modeling, not simulated emotions
• For Whom: Anyone who wants AI conversations that feel more natural without sacrificing honesty