Agent drift

[younessouilmi]

Hi guys, I've built a flowise flow featuring an OpenAI function agent with instructions to perform specific tasks and an output with precise content and format. I call the flow via the API to perform a repetitive task with no previous message memory: I use a buffer window memory with k=0.
The agent works well but regularly drifts from desired behavior after few 10s cycles; to be noted that once it deviates from the desired behavior it tends to repeat the same deviation multiple times in row.

I suspect it is doing some kind of "continuous learning". My only cure for now is manually prompting the flow via the UI/ deleting message history via the UI.

is their a way to prevent the drift from happening or a programmatic action to reset the flow?

Thanks in advance for your help!