Enhancing ChatAgent Memory Flexibility with Memory Callback Parameter

[coolbeevip]

ChatAgent can write to memory, which helps developers with business logic, but it needs more flexibility. I propose adding a memory_callback parameter to the step method to decide if a Q&A pair should be saved based on the input and output messages.

We should also move the input_message method to run after the model's result, allowing us to use the memory_callback return value to determine if it should be saved. This change may impact current memory recall logic, and we need to include an input_message to a copy of self.memory.get_context().

Here's a more specific example

def my_memory_callback(input_message: BaseMessage, output_message: Optional[BaseMessage]) -> bool:
    """
    A callback function to determine if messages should be recorded to memory.
                Takes input and output messages as parameters and returns a boolean.
                If True, messages are recorded. If None, default recording behavior applies 
    """

agent = ChatAgent(
    system_message="xxx",
    model=self.model_backend,
    message_window_size=10    
)

response = agent.step(prompt, memory_callback=my_memory_callback)

[coolbeevip]

After carefully reviewing the code, I've decided to abandon the previous idea of using a callback. Implementing this functionality within the step method would overly complicate the logic, making it difficult to maintain.

The underlying reason is that I want to have precise control over the tokens input into the large model. Is this achievable? For example:

1. Support List[BaseMessage] type for the input_message parameter in the step method.
2. Include all messages from the current round of conversation in the parameters returned by the step method.
3. I clear the internal memory using the reset method before each call to the step.

This way, I can manage all the tokens myself outside of the ChatAgent, and I can decide which tokens to pass in when calling the step method.