I swear it’s more like a human than AI sometimes. Like it has times when it’s straight up really, really dumb and useless and it’s so frustrating. Yeah, it forgets what it’s doing, or it says it did things it didn’t do.
I got something the same, Cursor forget what are original objectives, architecture we are doing, always creating new tools, routines, sub-routines even I have let Cursor know about the plans.
Additionally, when have a break during the day then came back to continue. It happen like what you got a new guy come to work on different way from previous and most of the time make it messy than before until you told Cursor to get back on track. I much worst than earlier versions.
Yes, it is a pain in the neck, is like programing with a teamate who has alzheimer. You are correcting some errors, and then it forgets and leaves part of the solution, but then it “remembers” what it was doing, but it changes the apporach and corrupts the code.
Today it made me waste like 6 hours of a program that it was running well and I asked to change the name in the interphase and it broke all the progress.
Why is everyone trying to call this a hallucination, or a context length issue? Hallucinations are random content generation when under stress or lack of idea what to do next. It’s not the same outcome every time like we’re seeing. It’s not context window size either, as sometimes it will occur on first command in a fresh chat.
It’s a bug. Somewhere, there is a turn point in the conversation or the act of scanning certain files, etc — somehow this action triggers a bug and context is flushed and it defaults to summary mode. What the trigger is, is the issue at hand.
Same issue here. I never spent this quickly 500 requests and I swear most of it came from it randomly forgetting what it was doing and reaching that “I did 25 calls; I will stop now, should I continue” point. This is a bug, not hallucination.
I’m going through the exact same issue, and honestly, sometimes it feels like it’s all part of a strategy to increase revenue — the more nonsense Cursor generates, the more requests we burn through, and the more we end up paying. It’s frustrating, especially when you give clear instructions and it still veers off into irrelevant code or gets stuck endlessly scanning. Add to that the constant disconnects and outdated web search results, and productivity just tanks. At this point, it’s hard not to question whether it’s worth sticking with it.
Dude! if you can’t control it, don’t smoke it!!! It might be a hallucination if all of us reporting the issue were at the same place. Don’t U think?
@MoacirParticular I’m starting to come to that same conclusion, I could be on a really long chat on Claude.ai and ask what the instructions or list of objectives and it would tell me instantly, whereas with Cursor it gets lost every quickly. Cusor does need to get their act together quickly because they’re not the only ones out there and I’ve tried their Competitors which is good but lack some of the more advanced LLM once they add them, then Curosr is in trouble.
Update: I’ve unsubscribed and am going with Roo Code and Replit. I’m at my wits’ end with Cusor. I’ve built a basic cross-sell Shopify app for a client and thought let’s use a working project to see if Cursor could enhance it by providing a few more options to the Merchants. That was a stupid mistake, the task was basically. I created 123 and I wanted Cursor to go 456, but I got 54??? Broken. Game Over I’m done!!