While working in Cursor, I observed that the agent appears to return output unrelated to my current project, as if another user’s session or context is being mixed into mine. Specifically, I was working on an AI-related function and asked the Cursor agent how to use that function. However, the response I received was about a completely different project involving a file converter, which is not part of my workspace or current task. This raises concerns about session isolation and potential cross-session data leakage.
Steps to Reproduce
Use the Cursor agent and ask a question about how to use the AI function in my project. Observe the response generated by the agent. The agent returns an answer unrelated to my project, referencing a different project (file converter).
Hey, thanks for the detailed report and the screenshot. I’ll start with the main point, this is not another user’s session leaking. Sessions are isolated between accounts, and you did not get any other user’s data.
If you look closely at the screenshot, the file-converter project is on your own drive d:\project\java\file-converter-service, the agent reply is in your language, and the panel shows your own chat tabs and the Undo All and Keep All controls for files in your workspace. So the agent pulled context from another one of your chats or an earlier task in the same workspace, not from someone else’s session. This is known behavior where context can bleed between chats, and it’s being worked on.
What helps in practice:
Start a new chat for each unrelated task, don’t reuse one long chat across different projects, so context won’t mix.
Update Cursor to the latest version, you’re on 3.4.17 which is a bit behind.
If you see this again in a fresh chat, please share the Request ID for that reply right top corner of the chat > Copy Request ID and we’ll dig in.
I understand your explanation, but what still concerns me is that I never asked the agent to create or work on any file-converter project at all.
In this case, I was only asking about an AI-related function in my current project. The file-converter context appeared unexpectedly, and I have not referenced or worked on that project in this chat session. That is why it felt like unrelated context was being introduced without any trigger from my side.
Could you help clarify how the agent decides to pull in context from other chats or tasks? I want to understand how to prevent this from happening, especially when working on completely unrelated projects.
Good question, let me explain how it works. The agent doesn’t deliberately decide to pull in another project, there’s no conscious choice. In a single chat, the model can access a few context sources: open files, the codebase index for the current workspace, and the chat history itself. When one workspace or a long session includes work on different tasks, the model can sometimes latch onto an unrelated piece of context, even if you didn’t mention it in your current message. This is known model behavior, not data being shared between accounts, sessions are isolated.
What helps reduce it in practice:
Open the specific project folder as the workspace, not a parent folder. In your screenshot the path is d:\project\java\file-converter-service, it looks like you opened d:\project\java\ with multiple projects inside. The narrower the scope, the less neighboring code ends up in the index.
Start a new chat for each separate task. Don’t reuse one long chat across projects. The longer the history, the higher the chance of mixing.
Update Cursor to the latest version. You’re on 3.4.17.
If it happens again in a fresh, clean chat, send the Request ID for that answer (top-right corner of the chat > Copy Request ID) and we can look at the specific case.
