Emotion-Aware Long-Term Conversational Memory

This repository presents an emotion-aware long-term conversational memory framework for Large Language Model (LLM)–based dialogue agents. The system enhances traditional Retrieval-Augmented Generation (RAG) by integrating emotional context into memory representation and retrieval, enabling more human-like, emotionally consistent, and personalized conversations over extended interactions.

---

🚀 Motivation

While modern LLMs excel at short-term dialogue, they struggle with:

• Long-term memory retention across sessions
• Emotional continuity and consistency
• Contextually appropriate recall of past interactions

Human memory is strongly influenced by emotion. Inspired by Mood-Dependent Memory theory, this project introduces emotion-aware memory retrieval, allowing conversational agents to recall information not only based on what was said but also how it felt.

---

🧠 Key Features

• Emotion Modeling with Plutchik’s Framework
  Extracts an 8-dimensional emotion vector (joy, acceptance, fear, surprise, sadness, disgust, anger, anticipation), with intensity scores from 1–10.

• Emotion-Aware Memory Retrieval
  Combines semantic similarity and emotional alignment using cosine similarity and Euclidean distance.

• Long-Term Memory Support
  Designed for multi-session conversations with persistent memory storage and retrieval.

• LLM-Based Emotion Extraction
  Uses LLaMA-3.3-70B-Versatile with structured prompts to generate stable emotion embeddings.

• Reflective Memory Design (Inspired by RMM)
  Supports flexible memory granularity and improved retrieval quality over time.

---

🏗️ System Architecture (High-Level)

1. User Query Processing

   • Extract semantic embedding
   • Extract emotion vector

2. Memory Retrieval

   • Semantic similarity (text embeddings)
   • Emotional similarity (emotion vectors)
   • Weighted joint scoring

3. Response Generation

   • Retrieved memory + current context
   • Emotionally aligned response generation

---

📐 Emotion Representation

Each user query is mapped to an emotion vector:

emotion_q ∈ ℝ⁸
emotion_q = [joy, acceptance, fear, surprise, sadness, disgust, anger, anticipation]

Where each dimension is an intensity score from 1 to 10.

---

⚙️ Memory Retrieval Algorithm

1. Compute semantic similarity: $$sim_semantic = cosine\_similarity(embedding_q, embedding_m)$$
2. Compute emotional similarity: $$dist_emotion = euclidean\_distance(emotion_q, emotion_m) sim_emotion = 1 / (1 + dist_emotion)$$
3. Combine scores: $$sim_total = α * sim_semantic + (1 - α) * sim_emotion$$ Where α is a tunable parameter (0 ≤ α ≤ 1).
4. Retrieve top-K memories based on sim_total.

---

SYSTEM PROMPT

You are a mental health counselor. Your task is to provide thoughtful and offer constructive advice or encouragement when appropriate. Keep replies to one line.

USER PROMPT

You are a compassionate mental health counselor.

Your role:

• Respond with empathy and emotional validation
• Gently reflect recurring emotional patterns without sounding clinical
• Offer practical, non-overwhelming coping suggestions
• Never mention internal data, scores, timestamps, or memory systems

Current user concern: {query}

Relevant memories from previous conversations: {memory_block}

Guidelines for using memory:

• If similar feelings occurred recently, acknowledge recurrence gently
• Use soft temporal language (e.g., "recently", "before", "this has come up again")
• Do NOT mention exact dates or frequency
• Do NOT quote past messages verbatim
• Integrate memory naturally into reflection

Response structure:

1. Validate the current feeling
2. Reflect any recurring emotional pattern if present
3. Offer one grounding or coping suggestion relevant to the concern

Tone:

• Warm, calm, non-judgmental
• Human and supportive, not clinical
• Avoid advice overload