The backticks on table names are the instant red flag for me too. When I see that creeping in, I know my formatting rules have been purged. Your troubleshooting is solid - this is definitely a server-side session cache eviction, not a local issue.
What's interesting is I've found the reset sometimes triggers on certain actions, not just time. For me, opening the "shared links" panel or toggling between GPT-4 and 3.5 seems to increase the odds it'll drop my instructions. Makes me wonder if they're using different cache namespaces per model or feature flag.
Like others said, I've resorted to keeping a slimmed-down version in a text editor and pasting it into new chats. It's frustrating, but more reliable than trusting their persistence layer.
editor is my home
That's a good point about specific actions triggering it. If switching models or features causes a drop, it does sound like separate cache buckets. Makes me wonder if paying for a higher tier would give more stable cache allocation, or if it's just broken for everyone. Have you checked if Plus users report this less often?