Multi-Turn Conversations with Claude in Python

Build multi-turn chat with Claude in Python. Manage conversation history, control context length, and implement memory patterns with the Anthropic SDK.

Since the Claude API is stateless, building a chat requires storing and sending the full conversation history each turn.

Simple chat loop

import anthropic

client = anthropic.Anthropic()

def chat():
    messages = []
    print("Claude Chat — type 'quit' to exit
")

    while True:
        user_input = input("You: ").strip()
        if user_input.lower() == "quit":
            break

        messages.append({"role": "user", "content": user_input})

        response = client.messages.create(
            model="claude-sonnet-4-6",
            max_tokens=1024,
            system="You are a helpful assistant. Be concise.",
            messages=messages
        )

        assistant_reply = response.content[0].text
        messages.append({"role": "assistant", "content": assistant_reply})
        print(f"
Claude: {assistant_reply}
")

chat()

Track tokens and trim history

MAX_CONTEXT_TOKENS = 150_000

def trim_history(messages: list, max_tokens: int = MAX_CONTEXT_TOKENS) -> list:
    """Remove oldest message pairs until estimated token count is under budget."""
    # Rough estimate: 1 token ≈ 4 chars
    while len(messages) > 2:
        total_chars = sum(len(str(m["content"])) for m in messages)
        if total_chars / 4 < max_tokens:
            break
        # Drop the oldest user + assistant pair (keep first message for context)
        messages = [messages[0]] + messages[3:]
    return messages

messages = []
cumulative_tokens = 0

def chat_with_trimming(user_input: str) -> str:
    global messages, cumulative_tokens
    messages.append({"role": "user", "content": user_input})
    messages = trim_history(messages)

    response = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=1024,
        messages=messages
    )

    reply = response.content[0].text
    messages.append({"role": "assistant", "content": reply})
    cumulative_tokens += response.usage.input_tokens + response.usage.output_tokens
    return reply

Inject retrieved context per turn (RAG)

def chat_with_rag(user_input: str, context_docs: list[str]) -> str:
    context = "

".join(context_docs)
    augmented_message = f"Context:
{context}

Question: {user_input}"

    messages.append({"role": "user", "content": augmented_message})
    response = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=1024,
        messages=messages
    )
    reply = response.content[0].text
    messages.append({"role": "assistant", "content": reply})
    return reply

For prompt caching on repeated system prompts, see the prompt caching Python example. For the Node.js equivalent, see the Node.js quickstart.

Frequently asked questions

Free tools

More examples