You ask Cursor to look into a bug. It starts to examine the issue, then starts to make changes, then it goes on a file scanning spree. Then it hits the 25 tool call limit. You press ‘continue’. It continues scanning, but it has completely forgotten what it was doing, instead after the scans it just provides a summary of how “x” feature works, and leaves it’s previous edits unfinished. This happens all the time now. This is on a fresh composer session.
Does hitting the tool call limit somehow flush its context completely now? I’ve never experienced this before, it would usually be able to continue and work on the problem it started, but now it’s like a whole new agent has stepped in without any idea what it was supposed to be doing, it’s very frustrating.
I am experiencing the same behavior. In addition, that it starts to (what feels like) randomly generate code and/or give out terminal commands to run. After which the whole projects is being disrupted. It seems like it completely lost any sign of context of the project.
I got all of that with the latest chat update that was introduced. Each time I open Cursor, it shows me a ‘New chat’, instead of continuing where it was left off last time. Could that be the issue of context lost?
This is the result of small context window. I experience the same everyday before I started to use other IDE.
Imagine that LLMs are as smart and have good memory as dolphins, but Cursor shortened their memory to 7 seconds, just like a goldfish… Goldfish will really keep eating because they forget that they have eaten, and then they will eat themselves to death
Yes, I recently encountered this situation, Claude-3.7-sonnet would read too many code files and completely forget the problem it was going to solve. After a long wait, it gave a summary of the code it read instead of solving the problem. This happened with a Pro subscription, and I don’t understand why Cursor gave such a short context limit.
I also encountered this situation. While troubleshooting, I found myself repeatedly checking the file structure even though I had just reviewed it in the previous conversation. After examining the project structure, I completely forgot the original problem I was trying to solve, then simply summarized the project and ended the conversation. This issue only occurs with Claude3.7, and I’ve since stopped using it.
This is happening to me as well. Agent mode is currently making project work incredibly difficult. Yesterday, I experienced a severe issue where the agent would jump between current and old task, causing a complete loss of coding context. Worse, it would randomly grab code from past conversations and start applying those changes to my current work. This forced me into a constant cycle of stopping the agent, salvaging correct code, and undoing unwanted modifications. My productivity was practically nonexistent. This problem is happening with every model I use, and starting a new chat doesn’t fix it.
And if you click to ‘continue’ I always always get a connection problem ‘check internet connection or vpn’. It has nothing to do with my connection. This has made being productive with Cursor impossible.
I’m the same. I spent an extra $20 last month over the 500 allocated request, and 90% of that spend has been fixing issues where Cusor goes on random journeys into the never-never that’s after giving specific instructions and as the OP stated, gets stuck scanning the Codebase. That’s on top of the constant disconnection issues. I find I get better results using Claude and Gemini standalone services. Another thing I find is that Cursor uses outdated information when you tell it to search the web, and returns information that’s not up to date. At this rate I might need to pull the plug on Cursor if things don’t change soon, because this is becoming really unproductive, there are too many issues with it.
I’m having similar issues, it sometimes gets carried away with trying to read extra files etc and doesn’t actually get to solving the problem, even for relatively simple problems when you have attached the context needed - please fix!
I think we are at the point, where the number of users is overloading the services forwarded to by Cursor AI and thus not getting the computational bandwidth to do the even basic stuff.
In my case it is now 5-10 minutes into extracting a single simple class out of main.cpp!
I am reporting the same issue. Not only it forgets what it’s working on, but it stops completely from working altogether. I am using “gemini-2.5-pro-exp-03-25” and it constantly loses connection. It has become unusable.
Update: I had to create a new chat to have it start moving. Once new chat created, I had to refer Cursor to my TODO.md to pick up lists of tasks from an outstanding task list. Now question: why wouldn’t Cursor work off existing chat session? Was it getting too big? If ‘yes’ is there a logic to implement for Cursor to trigger a ‘migration’ process (summarize existing chat, move to new chat) process and a warning alert, so we, mere mortals would understand what is happening?
I struggled for hours trying to implement a simple feature, Cursor constantly disconnects “check your network or vpn” nonsense message, constantly forgetting what it’s doing and just writing up a summary instead of implementing features or fixing bugs.
I switched to Roo Code and made more progress in an hour with Roo than the last 8 hours with Cursor.
THere’s something really wrong going on in Cursor’s backend, it’s just not productive at all right now.
