One of my colleagues created a really useful Cursor automation that i was eager to copy for my own team. However, even after I updated all the connections and the instructions, it still kept checking her team’s code, despite me making sure that nothing in the instructions said to do that, and I deleted all memories.
I finally created a new automation with the exact same information I had changed the copied-and-edited version to use. Critically, I created it without copying it. It worked perfectly from the first run.
Later, I updated the instructions of this new automation to message a different team chat. But it never actually posted to that team chat – it kept using the one from the previous version of the instructions. I was pulling my hair out trying to figure out what was going wrong. Again, I deleted the memories and tried re-running.
Finally, I remembered the copying issue. So, I created another brand-new automation, which once again, had the exact same instructions as the first one. And once again, the brand-new automation worked perfectly the first time, where the copied-and-edited version did not.
It seems pretty apparent to me that Cursor automations are preserving memories even when the user tells it to delete them.
Steps to Reproduce
Create Cursor automation in web interface
Copy the Cursor automation
Update the instructions. Notice that not all of the new instructions are followed.
Delete the memory for the automation, and save again. Notice that the same issues persist.
Rewrite the instructions after deleting the memory. Notice that the same issues persist.
Create a brand-new automation with the exact same instructions. Notice that it works perfectly on the first try.
Update the instructions for the new automation. Notice that the new instructions are not followed.
Delete the memory of the new automation and run it again. Notice that is still ignoring the new instructions.
Rewite the instructions for the automation and run it again. Notice that is still following the original version of the instructions.
Once again, create a brand-new automation with the same instructions. Notice that the instructions are, once again, followed perfectly on the first run.
Expected Behavior
Cursor automations should follow their instructions. If the instructions are updated, the automation should follow the new ones.
Operating System
Windows 10/11
Version Information
Cursor Cloud Agent
For AI issues: which model did you use?
I tried Codex 5.3 High, Sonnet 4.6 High, and Opus 4.6 High
Thanks for the detailed report and reproduction steps. This is a known pattern, and your observation about new automations working while edited ones don’t is an important clue.
I attempted to reproduce this by creating an automation with one set of instructions, running it, updating the instructions, and running it again. In my test, the updated instructions were correctly picked up on the second run. This suggests the issue may depend on how the automation is triggered or how quickly it’s re-run after editing.
A few questions that would help narrow this down for our engineering team:
What trigger type are your automations using? Scheduled (cron), GitHub events (push/PR), Slack, or webhook?
When you re-ran after editing, did you use the “Test Run” button, or wait for the next scheduled/triggered run?
How quickly after saving the updated instructions did the automation run again? Seconds, minutes, or longer?
When you duplicated your colleague’s automation, did you edit the instructions before the first run, or did you run it once first and then edit?
In the meantime, the workaround you’ve found (creating a brand-new automation instead of editing) is the right approach.
Hi, @mohitjain! Thank you for the reply. I’ll answer as best I can.
My automation uses two triggers. One is a weekly cron job that fires off on Friday at 9:00 Eastern US time. I’ve actually had no problems with this one, but that’s probably because each week, I’ve ended up creating a new automation before the scheduled time and resolved the issues. The other trigger is in Slack, where I mention a specific phrase in a specific channel. That’s the one I’ve had the most trouble with.
My test flow after updates has been to immediately use “Test Run,” and once that seems to be working, trying to trigger the automation from Slack as described in #1.
I generally test within minutes after creating/updating the automation. Sometimes it’s longer, but I can’t think of a time when I’ve run a test more than ten to fifteen minutes after an update.
When I duplicated my colleague’s automation, I updated the instructions first.
Thanks again for your help and please let me know if I can provide more information!
The Slack trigger being the problematic one narrows things down. One follow-up to make sure I’m reading this right: when you use “Test Run” after updating instructions, does the test run correctly follow the new instructions? For example, if you changed “post to Channel A” to “post to Channel B,” does the Test Run actually post to Channel B, and then only the Slack-triggered run reverts to Channel A?
Or does Test Run also follow the old instructions, and “seems to be working” means it runs without errors rather than following the updated instructions specifically?
This distinction would help confirm whether the issue is isolated to the Slack trigger path.
To answer your question: when I use “Test Run” after updating instructions, no, the test run does not correctly follow the new instructions. This actually led to an embarrassing problem where I changed “post to Channel A” to “send me a DM,” but the test runs kept posting to Channel A.
So no, I don’t believe this is isolated to the Slack trigger path.
Thank you for confirming that. This is an important detail: the fact that Test Run also follows the old instructions rules out the trigger type as the cause.
To help our engineering team dig deeper, two things would be helpful:
The automation URL from your browser’s address bar when you’re viewing one of the affected automations (it will look something like cursor.com/automations/``...). This lets us inspect the configuration directly on our end.
After you edit instructions and save, if you reload the page, do the new instructions appear correctly in the editor? This would help us determine whether the issue is with saving the instructions or with the automation not picking them up at runtime.
Our team is actively investigating this. Appreciate your patience and the detailed answers throughout this thread.