Since 1.2.4, the AI in Cursor has started to lose track of the current state of progress within a session. Specifically:
The AI loops through the active context documentation in nearly every response, even when it has just generated new code or completed a step.
It repeatedly restates or re-reads the project documentation, even when the session is uninterrupted.
The AI seems to “forget” progress or context between responses, requiring me to re-explain or re-confirm the current state.
This behavior is new; previously, the AI maintained context and only reviewed documentation at the start of a session or after a reset.
Steps to Reproduce
Start a session and provide project documentation and rules.
Work through a multi-step implementation (e.g., incremental code changes).
Observe that the AI repeatedly reviews documentation and loses track of progress, even within a single uninterrupted session.
Actual Behavior:
The AI acts as if it is stateless between every response and even seems to forget the current state of progress during a response following code generation, which is confusing and scary.
Expected Behavior
The AI should maintain context within a session, only reviewing documentation at the start or after a reset. It should not repeat documentation reviews or lose track of progress during or after every response.
Operating System
MacOS
Current Cursor Version (Menu → About Cursor → Copy)
I have not made any recent changes to my project rules or operating procedures. This issue only started recently, suggesting it may be related to a Cursor platform update.
I’ve noticed that Cursor has recently started adding features to automate the creation of rules and procedures. I have manually created comprehensive rules and procedures already, and maintain a memory-bank (Cursor Memory Bank · GitHub). There might be a conflict between the automated and manually-created rules and procedures that sends the AI in a loop that makes it appear to be losing context and repeatedly reviewing parts of the documentation. I had to switch to Auto because Claude 4.0 is completely unworkable, constantly just reviewing and re-reviewing no matter what I tell it to do.
This issue is not fixed as of 11:14am EST July 25, 2025. The quality of content generated by the AI’s are currently very low. I now switched to Claude 3.5 and it gets stuck in an infinite loop reviewing code. Here is an example of what that looks like right now:
Let me continue searching for the AJAX handler registration in [redacted] to ensure we have complete information before implementing the interception.
Ran tool
Let me continue searching for the AJAX handler registration in [redacted] to ensure we have complete information before implementing the interception.
Ran tool
Let me continue searching for the AJAX handler registration in [redacted] to ensure we have complete information before implementing the interception.
Ran tool
Let me continue searching for the AJAX handler registration in [redacted] to ensure we have complete information before implementing the interception.
Ran tool
Let me continue searching for the AJAX handler registration in [redacted] to ensure we have complete information before implementing the interception.
Ran tool
Let me continue searching for the AJAX handler registration in [redacted] to ensure we have complete information before implementing the interception.
Ran tool
Let me continue searching for the AJAX handler registration in [redacted] to ensure we have complete information before implementing the interception.
Ran tool
Let me continue searching for the AJAX handler registration in [redacted] to ensure we have complete information before implementing the interception.
Ran tool
Let me continue searching for the AJAX handler registration in [redacted] to ensure we have complete information before implementing the interception.
Ran tool
Let me continue searching for the AJAX handler registration in [redacted] to ensure we have complete information before implementing the interception.
Ran tool
Let me continue searching for the AJAX handler registration in [redacted] to ensure we have complete information before implementing the interception.
Ran tool
Let me continue searching for the AJAX handler registration in [redacted] to ensure we have complete information before implementing the interception.
Ran tool
Let me continue searching for the AJAX handler registration in [redacted] to ensure we have complete information before implementing the interception.
Ran tool