What's the correct way adding thread memory into a supervisor agent? How to also reduce number of messages in the conversation?

[arielbeckjit]

Basically - i don't need the checkpoint mechanism, it saves too much data and can cause overload on the DB.

supervisor_workflow = create_supervisor(
    [a1, a2, a3],
    model=model,
    state_schema=MyState,
    prompt=(
        "the prompt"
    ),
)

memory = DynamoDBSaver(config=config, deploy=False)
supervisor_agent = supervisor_workflow.compile(checkpointer=saver)


async def route_query(
    user_message: str, user: dict,
    thread_id: Optional[str] = None
):

    """
    Routes a user query to the appropriate agent using the supervisor workflow.
    """

    # Prepare the input for the supervisor
    graph_input = initial_state()
    graph_input["messages"] = [HumanMessage(content=user_message)]

    # Prepare the context data
    config: RunnableConfig = {
        "configurable": {"user_details": user},
        "callbacks": [get_observability_callback()],
        "thread_id": thread_id,
    }

    # Process the query through the supervisor workflow using stream_mode for more control
    async for chunk in supervisor_agent.astream(graph_input, config):
        # Print debug information about the chunk
        print(f"\n\n------ chunk: {chunk}\n\n")

the main problem is that the checkpointer saves tons of checkpoints data, it will get expensive and i dont need checkpoints, just message history (human / AI - that's it)

The other issue is that items start to build up - an item can suddenly have 100 messages (it doesn't save individual messages per checkpoint)

Do i need to build it on my own? load it and save it at the end of the graph?
how to address this?