the composer cant even take 2 prompts without forgetting what id just did.
over and over again deleting and creating the same files.
cursor is completly unusable now
Hi! I’m in the same situation for the past few days, and honestly, it’s really frustrating. Cursor can’t seem to keep the context of files, which makes it impossible to move forward with my project. I’ve tried leaving it for a couple of hours to see if the model “recovers,” but so far, no luck. I’m completely stuck and unable to continue.
I really hope this gets fixed soon because working like this is extremely challenging. Is anyone else experiencing the same issue? Any solution from CURSOR?
no they just tell you they fixed everything and tell you its a problem on your site but thats not true.
cursor gets more and more unusable because they are using a old version of VS Code and they encountered a lot of issues since they dropped the update with the composer agent
If you are using the normal composer, make sure you @ the files to keep the context available to the AI.
The agent may sometimes need prompting to look at a file explicitly, as sometimes it assumes it knows all the context when it doesn’t! We’re always working to try and improve this, and agent is our newest feature so is still be working on, but adding a line like “Make sure you look at the contents of a file before you edit it every time!” in your prompt should ensure it has all the context it needs to work well!
Additionally, when composer chat’s get too long, the LLM gets worse at keeping all the context from the whole chat, so it’s best to start new chat sessions every so often, especially when working on a new feature or areas of your code.
I’ve wasted hours what used to be simpler before. It just keeps dumbing up if there’s such a word
My experience is similar, totally frustrating. Unable to hold or execute context and instruction. Something completely messed up. It deletes good code compulsively, no one asked for it, but it deletes it. It fixes and breaks where there is working stable code, all without asking. I spent days trying to protect the finished pieces of code from his erroneous interventions. Terribly frustrating. I don’t know what happened, but something is not quite right now. It would be nice if there was a tool for this purpose, to indicate which parts should not be touched.
Hey, you can set this as a strict rule so that the model doesn’t delete your code. You can also create marker tags in the comments to indicate that the marked code should never be touched. This should work.
Although Cursor has made some progress with the 0.445 update, which is commendable, to be honest, it still doesn’t perform as well as the 0.42 version. While it sometimes delivers impressive results, there are times when it can be quite disappointing. In some cases, you may need to rebuild the code repeatedly, possibly due to the lack of support for long context, or perhaps I just don’t seem to be a good fit with Cursor. In contrast, Windsurf has been performing quite well recently, while Cursor’s performance remains inconsistent and unstable.