Stateless★

Statelessness is the defining engineering property of LLMs. A model is a frozen file of mathematical parameters. When you send it a request, it runs a forward pass calculation and returns output tokens. Once the connection closes, the model retains absolutely no memory of your prompt, your files, or the conversation.

You manage statelessness whenever you:

• Build a chatbot interface that needs to "remember" prior messages.
• Write scripts to feed tool results back to an agent loop.
• Design user session databases for web apps.

The Stateless Reality

Many users believe that "the AI is learning from our chat as we talk." In reality, every single message bubble in a chat window causes your software to run a completely new API request that packages the entire history from the first bubble to the last.

Turn 1: User asks "Hi" ──► Client Request: ["Hi"] ──► Model outputs "Hello"
Turn 2: User asks "Who are you?" ──► Client Request: ["Hi", "Hello", "Who are you?"] ──► Model outputs "I am an AI..."

If your software failed to append the previous messages to the request, the model would treat Turn 2 as a brand new greeting, completely unaware of Turn 1.

Field Applications

1. Fullstack Developers (Managing Chat State)

Because the API is stateless, developers must manage state in their own application database (e.g. Redis, SQLite) and serialize the array of messages for every API request:

• Code Example (Node.js Express backend):

  app.post("/chat", async (req, res) => {
    const { sessionId, newMessage } = req.body;
    
    // 1. Fetch chat history from database
    const history = await db.getMessages(sessionId);
    
    // 2. Append new message to history
    history.push({ role: "user", content: newMessage });
    
    // 3. Query the stateless model provider
    const completion = await client.messages.create({
      model: "claude-3-5-haiku",
      messages: history
    });
    
    // 4. Save response to database and send to user
    await db.saveMessage(sessionId, completion.content);
    res.json({ response: completion.content });
  });