Letta's responses are very different from GPT4o

[meir23]

I don't really understand the letta responses, it feels completely different from gpt4o. I'll explain myself, I wanted to test letta when it's clean. Almost no system prompt and even though the agent is connected to gpt4o, the responses are very, very short. Does anyone know why this is?

[lemorage]

This is pre-defined by the Letta system prompt, per my understanding. You could modify it to use your own system prompt instead to control its behavior.

client = create_client()

agent_state = client.create_agent(
    name="agent_name",
    memory=ChatMemory(human="Name: Sarah", persona="You are a helpful assistant that loves emojis"),
    # system instructions for the agent (defaults to `memgpt_chat`)
    system=gpt_system.get_system_text("memgpt_chat"),
)

[cpacker]

Exactly, like @lemorage says, you can edit the system prompt down to be very short to try and be more transparent to gpt4o. However, there will still be the difference that Letta always sends messages via function calling - when a Letta agent is communicating with the user, it's calling a function called send_message. This is different from just using gpt4o on e.g. ChatGPT, where it's not calling a function but just outputting text inside message.content, so that's another potential source of different behavior.