Multi-Turn conversations

Key Problem: No Message Storage

• The Anthropic API and Claude do not store any messages sent to them
• Each API request is independent with no memory of previous interactions
• Without intervention, follow-up messages have no context about earlier conversation

Solution: Manual Message History Management

• You must manually maintain a list of all messages in your code
• You must include the entire message history with every follow-up request
• This allows Claude to see the full conversation context

Message Structure

• Messages are stored as a list of dictionaries
• Each message has two fields:
• role: either "user" or "assistant"
• content: the text of the message
• Order matters—messages should reflect the chronological flow of conversation

Three Helper Functions for Conversation Management

add_user_message(messages, text)

• Takes a message list and user text
• Creates a user message dictionary and appends it to the list

add_assistant_message(messages, text)

• Takes a message list and assistant text
• Creates an assistant message dictionary and appends it to the list

chat(messages)

• Wrapper around the messages. create API call
• Takes the full message history as input
• Returns just the text content of the response

Conversation Flow Pattern

• Create empty messages list
• Add user message with initial question
• Call chat() with messages list
• Add returned assistant response to messages list
• Add follow-up user message to messages list
• Call chat() again with updated full history
• Repeat steps 4-6 for continued conversation

Takeaways

• The Claude API doesn't store messages—you must manually maintain the full conversation history in your code and include it with every request
• Messages are stored as a list of dictionaries with role ("user" or "assistant") and content fields, in chronological order
• Use helper functions to add user/assistant messages to the list and call the API, then always pass the complete updated history for follow-up requests
• The conversation pattern is: create empty list → add user message → call chat() → add assistant response → add next user message → call chat() again with full history