I was working on a complex task, but the AI didn’t implement it according to my requirements. I reverted all the changes using Git, discarded every modified file, deleted the entire chat, started a new Agent chat, and asked it to build the same feature from scratch.
However, it still implemented the exact same approach that I had already discarded.
Could someone explain why this happens? Is it because of Cursor’s cache, context, or token reuse?
I even explicitly told the Agent not to use any old cached context or previous data, and only then did it finally implement the feature according to my requirements.
Is there any way to completely clear the old cache or prevent the Agent from reusing previous context? I’ve been facing this issue repeatedly for the past month, and it’s becoming really frustrating because it wastes a lot of tokens and time.
Sometimes I clearly mention, “Don’t use any previous cached context,” but it still follows the old approach. In many cases, the only workaround I’ve found is switching to a different model.
Has anyone else experienced this? If there’s a better workaround or if anyone from the Cursor community or Cursor team can explain what’s happening, I’d really appreciate the help.
Hey, let’s clear this up. There’s no hidden server-side cache that mixes old chats into new ones. A new Agent chat starts with a clean conversation, and in Privacy Mode we don’t store conversation data on our side at all. What some call prompt or token caching is only a speed and cost optimization, it doesn’t change what the model outputs in a new chat.
So why do you keep getting the same approach?
Model determinism. With the same codebase and an almost identical request, the model will naturally converge on the same design. That’s why switching models “fixed” it, a different model has a different prior for the same task.
Local leftover artifacts. git discard or revert removes changes in tracked files, but it does not touch untracked files. The agent often leaves those behind: markdown plans or notes, auto-created rules, or memories. Those can pull a new chat back to the old solution.
How to really start from zero:
Run git status and git clean -nd dry-run to see any remaining untracked files like plans and notes from the previous agent, then clean them up.
Check Cursor Settings > Rules and see if there are any auto-created rules or memories that push you toward the old approach.
Most effective: state constraints positively. Instead of “don’t use the old cache”, write: “I tried Approach X, it didn’t work because reason. Implement it using Y.” That helps the model avoid rebuilding the same design.
Try that and tell me if it helped. If it still goes back to the same approach after cleaning untracked files and rules, let me know and we’ll dig deeper.