GPT-4.1 retains Memory Across Calls Despite Stateless

API
,
1

Hi all,
I’m currently using the OpenAI API with the latest SDK and GPT-4.1. Here’s a simplified version of my code:

new_prompt = ‘…xyz’

from openai import OpenAI

client = OpenAI(api_key = your_api_key)

response = client.responses.create(
model=“gpt-4.1”,
tools=[{“type”: “web_search_preview”}],
input= new_prompt)

According to the docs, API calls should be stateless — i.e., the model should not retain memory of previous prompts unless I explicitly include them in the messages list. However, I’m observing behavior that suggests the model is remembering previous interactions, even when I send an entirely new prompt.

I’ve verified that:

  • I’m not reusing the messages list.
  • I’m not using any wrapper like LangChain or memory-enabled classes.
    Is there any caching, tool memory, or known behavior in GPT-4.1 (especially when using tools like web_search_preview) that might explain this? Has anyone else encountered something similar?
2

Hi and welcome to the community!

It is most likely something with your implementation that’s causing this issue.

The short answer is: no :slight_smile:

What you can try is to perform your test via Curl to break down the requests to their most simple form.

2 Likes
3

Verify the value of new_prompt and make sure you aren’t passing any previous_response_id.