Python: GroupChatManager Lacks Chat History for Follow-up Questions

[MarekFabinski]

I am working with the GroupChatManager in Python, following the official documentation on Agent Orchestration with Group Chat.

My goal is to create a multi-turn conversation where agents can respond to follow-up questions by retaining the context of the preceding conversation.

Expected Behavior

• User asks an initial question: "Give me a list of 5 products from Contoso."
• The GroupChatManager selects the appropriate agent to answer.
• The selected agent answers: "Here are 5 products from Contoso: [Product A, Product B, Product C, Product D, Product E]."
• User asks a follow-up question: "Provide me with 2 more."
• The GroupChatManager understands the conversational context and re-engages the same agent.
• The agent answers by recalling the previous turn: "Here are two more products: [Product F, Product G]."

Actual Behavior
The GroupChatManager and the selected agent do not appear to retain the chat history. When the user asks a follow-up question ("Provide me with 2 more"), the agent fails to understand the context and asks for clarification, such as, "What products are you referring to?". The history of the conversation is lost.

What I've Tried
Managing the conversation in a while loop and attempting to pass the chat history back to the manager on each turn.
Investigating the use of Azure AI Foundry threads to maintain a persistent conversation state, but I could not find a clear implementation path for integrating this with the GroupChatManager.

Question
Is there a standard method for preserving and passing chat history to the GroupChatManager to enable contextual follow-up conversations? Could this be a potential bug, or am I missing a step in the implementation?

Any guidance on how to properly manage conversation history in this scenario would be greatly appreciated.

[lavinir]

Hi @MarekFabinski,

I was just looking at the same thing and had the same question. Did you find an answer for doing this using a full chat history ?

[MarekFabinski]

Hi @lavinir, I actually switched from using GroupChatOrchestration to MagenticOrchestration in my solution and figured out that you can use chat history object to keep the chat context and allow follow-up questions. It should be achievable for any orchestration pattern. Just make sure you manually add new messages to it after each loop and pass it to the chat manager along with the next task that you want your agents to work on.

[lavinir]

Thanks for the quick response @MarekFabinski!
I'm not sure I follow though. Invoking is done the the GroupChatOrchestration / MagenticOrchestration's invoke() method correct?
That method accepts a task as a string. How are you injecting the ChatHistory object again on the next turn ? Or am I missing something here.

[MarekFabinski]

Sorry @lavinir, I missed a crucial part in my answer. You are right, the invoke method accepts only string as task. For this I prepared a simple function that converts the Chat History object to string before passing it to task in invoke(). Here is sample of what I have:

def format_history_for_prompt(history: ChatHistory) -> str:
    """ Formats the chat history into a string for the prompt. """
    if not history.messages: return ""
    relevant_history = history.messages[-MAX_HISTORY_SIZE:]
    prompt_parts = [
        "Here is the conversation history so far. Respond to the latest user message based on this context.",
        "--- CONVERSATION HISTORY ---"]
    for msg in relevant_history:
        timestamp = msg.metadata.get('timestamp', datetime.now()).strftime('%Y-%m-%d %H:%M:%S')
        actor = "User" if msg.role == AuthorRole.USER else msg.name or "Agent"
        prompt_parts.append(f"[{timestamp}] {actor}: {msg.content}")
    prompt_parts.append("--- END OF HISTORY ---")
    return "\n".join(prompt_parts)

And then to invoke it:

chat_history = ChatHistory()
    while True:
        user_input = await get_user_input_async()
        if not user_input or user_input.lower() in ["exit", "quit"]: break

        chat_history.add_user_message(user_input)

        handle = await orchestration.invoke(task=format_history_for_prompt(chat_history), runtime=runtime)

Later you need to remember to capture all of the agents responses into Chat History object. Let me know if anything is still unclear.

[lavinir]

Thanks @MarekFabinski , appreciate the detailed response.
So basically you're working around the API limitation of accepting just a single input string by serializing the entire history as the input, got it.

I still think it's interesting that there isn't a native way to do this in the SDK itself, curious if this is by design.

[MarekFabinski]

@lavinir You're right, I did have to implement a workaround. My understanding of the Semantic Kernel documentation is that while multi-agent orchestration patterns do support human-in-the-loop interaction, this seems primarily intended for clarifying questions or notes within a single, ongoing task or input. It doesn't appear to be designed for continuous, chat-like experiences where users provide entirely new tasks or inputs sequentially.

I wouldn't be surprised if this functionality evolves given how frequently SK is updated. Alternatively, it's possible I've overlooked a more direct way to achieve this within the current framework. As I'm relatively new to SK and had to prototype something quickly for a client, I'm currently working with what I have, and my solution seems to be working well for now.