README.md

Context Compression Module (CCM) — Travel Agent

Make every token count. A plug-and-play context compression system that teaches AI agents what to remember, what to forget, and what to re-inject.

---

Overview

Modern AI agents break when conversations get long. They forget, lose track and recommend things the user already cancelled. — not because the model is bad, but because the context window fills up with noise.

This project builds a Context Compression Module (CCM) that sits between the agent and its growing history, keeping only what matters.

It is demonstrated through a Multi-City Travel Planning Agent that handles flights, hotels, restaurants, budgets, and user preferences across 20+ turn conversations.

---

How Memory Works

Three memory tiers feed the LLM — instead of dumping everything:

Tier	Storage	What it holds	Always sent?
🔴 Working Memory	JSON file	Allergies, budget, dates	✅ Yes
🟡 Episodic Memory	ChromaDB	Summarized decisions	RAG only
🔵 Archived Memory	ChromaDB	Compressed tool results	RAG only

---

How the Pipeline Works

Every user message goes through 4 steps before the LLM sees it:

User Message
     │
     ▼
1. MEMORY EXTRACTOR     → pull facts (allergy, budget, destination)
     │
     ▼
2. STALE DETECTOR       → mark cancelled info as invalid
     │
     ▼
3. RAG RETRIEVAL        → fetch only relevant memories
     │
     ▼
4. CONTEXT ASSEMBLER    → pack into 2,000 token budget
     │
     ▼
  LLM Agent (Llama 3.1 8B)

Result: Baseline sends ~11,200 tokens. CCM sends ~1,400 tokens. Same task. Better answers.

---

Prerequisites

Python 3.9+
A free Groq API key → https://console.groq.com

git clone https://github.com/mohit-bishnoii/Context-Compression-Module.git
cd Context-Compression-Module
python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install -r requirements.txt

Create a .env file:

GROQ_API_KEY=gsk_your_key_here

Verify everything works:

python test_setup.py

---

Run It

# Chat UI
python ui/app.py

# Evaluation (baseline vs CCM comparison)
python evaluation/run_evaluation.py

---

Key Results

Metric	Baseline	CCM
Tokens at turn 16	~11,200	~1,400
Allergy remembered	❌	✅
Budget tracked	❌	✅
Stale info removed	❌	✅
Response latency	8.2s	2.1s

---

Stack

What	Tool
LLM	Llama 3.1 8B via Groq (free)
Embeddings	all-MiniLM-L6-v2 (local)
Vector DB	ChromaDB (local)
UI	Gradio
Training needed	None

---

The deliverable is the CCM module — not the chat interface. Drop it in front of any agent that suffers from long conversations.