The current version of the auto mode has a significant drop in quality compared to the previous version. Now, when performing coding tasks, auto almost forgets what it has written, requiring multiple reminders, or even inputting the relevant files and code simultaneously, before it can recall it. Perhaps this is due to the introduction of Kimi 2.5? Or perhaps there are other reasons. At present, I find it hard to trust the output of auto.
We’ve had a few similar reports about auto mode context handling lately. To dig into your specific case, I need a bit more info:
Your Cursor version Help > About
A concrete example, one conversation where the quality dropped, plus the Request ID from that message right upper corner chat menu > Copy Request ID
Without a specific case to trace, it’s hard to tell what’s happening on the backend.
In the meantime, a couple of things that help with context loss:
Start new chats more often, especially before the conversation gets long. Summarization can drop important context in longer sessions.
Put key instructions in project rules .cursor/rules rather than in the chat. Those persist even after summarization.
Use @ mentions to pull in the files you need instead of relying on the model to remember them.
Regarding Kimi 2.5 specifically, auto mode picks the best model for each task, and the model mix does change over time. If you want to test whether a specific model works better for your case, you can select one manually instead of using auto.
Let me know if you can share a request ID and I’ll take a closer look.
I purchased the Cursor Pro annual plan on Oct 7, 2025. Since then, using a specific model has started prompting me to “pay more.”
Later, Cursor models such as Composer 2 and Premium were introduced. However, I still cannot use them without additional payment.
At the same time, the quality of the Auto mode has noticeably deteriorated: the agent does not retain context, ignores the specified working conditions, and continuously attempts to execute commands or scripts on its own, again disregarding those conditions.
I chose the annual subscription so I wouldn’t have to worry about extra costs for a year. But now I either have to pay more or go back to the “pre-AI era” and check and write everything manually.