Multi-Turn Conversations

Agents don't remember anything by default

Each run_sync call is independent, the agent has no memory of a previous call unless you give it one. This is the same statelessness you saw in the LangChain course: a model call is a pure function of the messages you send it, nothing is remembered server-side.

Carrying history forward

Every AgentRunResult exposes new_messages(), just the messages produced by that run, and all_messages(), the full history including whatever was passed in. Pass either into the next call's message_history= to continue the conversation.

from dotenv import load_dotenv
from pydantic_ai import Agent

load_dotenv()

agent = Agent("google:gemini-3.5-flash-lite")


def main() -> None:
    conversation = [
        "My favorite color is teal. Remember that.",
        "What is my favorite color?",
        "What did I just ask you?",
    ]

    history = []
    for user_input in conversation:
        result = agent.run_sync(user_input, message_history=history)
        print(f"User: {user_input}")
        print(f"Agent: {result.output}\n")
        history = result.all_messages()

    print(f"Total messages accumulated: {len(history)}")

message_history accepts ModelMessage objects, the same typed representation of a request/response pair the SDK builds internally, you're not hand-assembling role/content dicts the way you might with a raw chat completions API.

Building a conversation loop

Reassigning history = result.all_messages() each turn, rather than new_messages(), is what keeps the whole conversation, not just the latest exchange, in context for the next call. This is the direct equivalent of managing a growing list of HumanMessage/AIMessage objects yourself in LangChain, just with the accumulation done for you by all_messages().