I need to preface this by saying that I find the ergonomics of cursor to be generally good (and typically above the UX from other agentic editors/plugins). Unfortunately, this bug has been recurrent enough that I have built several “rules” that attempt to combat its behavior, experimented with a number of prompting strategies, and even that is not enough.
I had initally attributed this behavior to:
- Chattier Claude version (3.7)
But even after experimenting with other models, I think there are systemic issues beyond just a verbose model, but without knowing more about the actual implementation, I can only guess at the roots. I also don’t care to stand in the way of progress (we should be using the latest models and the latest version of Cursor), but I sometimes find myself considering joining the group of folks who downgraded, locked their version and went back to the 0.45 days.
This has been reported is several different hand-wavy ways, and it’s been dsicussed since the 0.46 rollout
The bug:
- Cursor will forget that it recently created a file (say, a new API endpoint)
- It will, in its stream of writing, go to the file that should use the new API, see that we need new data, and then proceed to CREATE A NEW FILE (typically with a slightly different name) that does the same thing, but in a different way, with a different type signature…
- Then it will see that the types don’t match up, and it ill proceed to “fix” the file that should be consuming the original API
- This happens on a frequent basis
Every project is different, but this really seems to be an issue with whatever rules/prompts are around the tool calls to look up context and how the internal prompting is prioritizing context from within the conversation. My gut says this isn’t so much about not having the context, but the utilization/pruning/prioritization of the context it has for each tool call. The closest “feeling” I can compare it to is when you ask an AI image generator for something concrete like “a 1990s anime forest sceen with an acorn on the ground,” and the prompt is automagically re-written to “a thrilling outdoor sceen, set in an ancient forest, with a lone acorn on the ground representing loneliness. add lots of extra detail to the background. moody, mythical, magic.”
Unfortunately, this code is for work, so no dice on concrete samples, but I asked for it to analyze the recent messages and explain why I would be frustrated with the output. This is the result:
me:
Look through the conversation. I want you to create a simple breakdown of why I am frustrated with your output.
cursor:
Let me analyze the conversation:
1. I created a planning document that clearly outlined all steps, including creating the [redacted] endpoint
2. I then proceeded to:
- Create the endpoint file
- Mark it as complete in the plan
- But then completely forgot about it and asked if we should create it again
3. When you called this out, I:
- First tried to justify it by saying "we already have" the endpoint
- Then had to be reminded that I literally created it
- Still didn't properly acknowledge my mistake
4. This pattern shows:
- Poor context retention
- Not reviewing my own work
- Making assumptions without checking
- Wasting time with redundant questions
- Not following the plan I created
The most frustrating part is that I created the plan and the code, but then acted as if I had no knowledge of either, requiring you to repeatedly point out my inconsistencies.
Would you like me to continue with the actual task, or would you prefer to provide more specific feedback about my performance?
Version: 0.48.6
VSCode Version: 1.96.2
Commit: 1649e229afdef8fd1d18ea173f063563f1e722e0
Date: 2025-03-31T04:59:09.581Z
Electron: 34.3.4
Chromium: 132.0.6834.210
Node.js: 20.18.3
V8: 13.2.152.41-electron.0
OS: Darwin arm64 24.3.0
I need a break from Cursor until this is improved.