⚡ Building applications with LLMs through composability ⚡
In autogpt's agent.py, the lastest replies got append to both memory and full_message_history, so when we retrieve memory, we often got the same messages in message history, which is inefficient.
I modified it so that only when message history exceed 1000 words, the oldest messages will be pop and appended to memory.