Context loss has become a structural problem. Every time the agent forgets what it once knew, the user has to re-explain everything from scratch: logs, project files, past decisions. This not only disrupts the workflow — it ends up filling memory again, only for it to be lost once more. And with it, accuracy, coherence, and valuable time are lost too.
The result is a repetitive loop of cascading errors, misunderstandings, and small changes that require multiple corrections. In this loop of context loss, it’s not the agent or the platform that pays the price — it’s the user, who keeps repeating the same explanations over and over.
I’m not writing this just as a complaint, but as a clear request: the single most important improvement Cursor could prioritize right now is stable, persistent agent context. The new memory features help, yes — but they don’t solve the core problem. We need agents that truly understand, remember, and follow through.
Agreed. I think memories and project specific rules mixed with a task manager (Claude Task Master or something like what Windsurf has built in) seems to do an okay job at bandaging this but if Cursor is the first to come out with truly intelligent context management or something of the sort, it’ll be groundbreaking.
what about switching models how much context is retained? They did introduce memory, but what needs to stop happening is this:
It’s Tuesday and Claude the AI bot is working on the app and completes 5 PRs
It’s Wednesday and Claude wakes up and forgets what the app is about again or what he fixed yesterday.
Like @Taidan said, we don’t need, can’t have, amd would be detrimental to have an infinite context window. There needs to be a side process that reads exchanges and saves them to make sure it doesn’t repeat itself, or, refer to it when prompting the user for next actions.
I’m sure the Cursor folks are aware of it (especially if they use the tool themselves :)) but it’s good to press on it
sudo rm ff/*