🎯 Core Goals
- Show simultaneously what the user sees (chat) vs. what’s actually sent (the growing bundle).
- Make visceral the fact that the ENTIRE conversation history is re-bundled and sent every turn.
- No system prompt or custom preprompt yet — pure conversation history only.
The LLM doesn’t “remember” your conversation. Every time you send a message, the app bundles the entire history — all previous messages — and sends it to the LLM as one big package. The LLM reads it all from scratch, every time.
👁️ Visuals & Interactives
The Growing Bundle
Every "Send" re-bundles the entire conversation history — and starts the LLM reading from scratch
Turn 0 / 4
👤 What You See (Chat UI)
Press Send to start...
📦 What's Actually Sent to the LLM
Bundle is empty...
↓ entire bundle ↓
🤖 LLM reads it all from scratch
↓
📝 Key Concepts
- The LLM Has No Memory: Each turn, the LLM starts from zero. It only knows what’s in the bundle sent right now.
- The App Builds the Bundle: The chat interface silently collects the entire conversation history and bundles it with the new message on every send.
- The Bundle Grows: Every turn adds two messages. By Turn 4, the LLM reads 7 messages at once — just to answer the latest one.
- Cost & Speed: Bigger bundles = more tokens processed = slower and more expensive. Many providers offer caching to lower costs
The LLM doesn’t “know” you — it re-reads the transcript every single time. Modern LLM services often include a MEMORY mechanism, it often is just a way to bundle your previous conversation into your current one somehow. We will discuss some common methods later.
🧠
QUIZ
When you send your 5th message in a chat, what does the LLM actually receive?
Only your 5th message — it remembers the rest from earlier
The entire conversation history bundled together and sent from scratch
A compressed summary of previous messages plus your new one