Cursor seems to hallucinate much faster now than in the past - did something change in the last 24h on cursors side?
Total experience w/ cursor: 14h per day for ~50 days.
I have learned to always create a new composer window when it starts to hallucinate in certain ways - this often can be a few hours of using it but can sometimes survive until even 6 hours of continuous use before it becomes unusable.
Since yesterday even the first messages now are hallucinating, where it will recreate an entire class/ script that it already know acknowledged existed.
My Q is to the cursor team - what is the best way now to use cursor with regards to providing context to the project?
I usuall would add a message to the new composer and tag all the files in my project (~60 scripts). And have it provide me it’s understanding of the project before I proceed. Now when I do that it seems to not understand anymore and hallucinates immediately on the first message.
Any ideas or advice for how to work around these limitations?
Limiting it to just one or 2 scripts of context is not enough because this is a highly complex project that requires a lot of context for any change.
Thanks in advance for your helpful constructive responses
Based on the docs, there are a few things you can try to maintain better context without tagging all files at once:
Use @Codebase (Ctrl+Enter) instead of manually tagging files - this is optimized for searching across many files and should give better results
Try using @Folders to reference entire directories instead of individual files Cursor – @Folders
Set up project rules in .cursor/rules to automatically provide consistent context Cursor – Rules for AI
The issue is likely related to token limits - when tagging too many files at once, you may be exceeding the context window (40k tokens for chat). This can cause Cursor to only use partial context, leading to hallucinations
For complex projects, I’d recommend starting with @Codebase searches and then adding specific files/folders as needed, rather than trying to tag everything upfront
So I decided to start testing using @codebase (and filtered to *.cs files only) and on the second attempt it just doesn’t work well.
"Let me search through the codebase for any subscribers… However, I don’t see the full codebase context for where these subscribers might be. Could you share any files that might be subscribing to "
What we need is a larger context window - I read that Sonnet 3.5 has a context window of 200k - so at what level is this limit of 40k tokens being put on it? Is that per prompt response or does that accumulate over time in the Composer instance?
I appreciate any answers on this one as it’s the biggest limitation of Cursor at the moment and I’m trying to understand is that imposed by Anthropic or by Cursor.
I would be more than happy to pay double for double the context window for example. Or 4x for 4x the context window.
Good news - we’ve recetly added a large context setting, which is available in the latest version. When enabled, it increases the maximum context window, so more of your files get sent to the AI in one go - keep in mind this will use more fast requests!