I’v noticed this several times but today it became evident that cursor manipulates model usage which is quite deceptive in my opinion. I’m a heavy user no doubt, sometimes spending up to 13hrs straight. I realized that at about 8-9hrs into my work the model will suddenly stop working, and forces me to open a new composer. The new composer works but unknown to me the model I was using (3.5 Sonnet) has been changed, but still shows as 3.5 Sonnet at the bottom. How do I know this? During my hrs of coding, I sometimes give instructions to draw diagrams for a visual representation of my prompts which the model gladly obliges to. But when I asked the model again to draw the same diagram, after 8-9hrs of work, I got the following response: “I apologize, but I don’t have the ability to draw or manipulate images directly.” So the question is “how can the same model that gladly drew my images earlier suddenly becomes dumb and not able to draw again?” It shows that the model was changed! I love Cursor but this behavior is manipulative. If anything needs to be changed please inform the user!
Context and inference is how information rich Claude is at the time of inference. Because of how Cursor feeds metadata into Claude as context, and has some degree of computer use capabilities enabled on Claude by the team, I highly suspect you’ve just found the limit to the tricks for LLMs that make them seem magical - it can only go so far with the context size of 200k tokens on Claude.
In this case, I imagine cursor team is giving Claude with some context on the project, metadata, notebooks, etc. and then your last X tokens (maybe 500 or so). The idea is that the model can still stay coherent on where you’re probably wanting to go next given even a limited amount of context, as long as that context gives it the tools needed to understand your code at the time of inference.
Here’s the solution: Are you asking Claude to take notes in a notes document? When you come back or find Claude losing context, go ahead and tell it to review the notes to recontextualize. You’ll see it should go right back to (or close to) what you remember in the first 8-9 hours.