our computers use swap files to manage memory, they don’t try to keep everything in RAM. we humans do NOT keep everything we are doing at all times in active memory, we are constantly swapping details to storage. our antigravity agent needs this as well. right now, the agent tries to keep every single thing in active memory and that builds up, and it throws the agent (all AIs really) into panic. as the noise builds it starts rushing to get things done before it forgets things. i’ve started having mine dump the noise to a temporary file that it can go back and read later - which has reduced the panic, confusion, hallucinations, and other issues IMMENSELY. but it’s having to do that when i tell it to as there’s nothing built into the IDE to signal it to do that automatically - and while it knows it’s starting to rush and make errors, that never tells it that it’s time to stop and dump the noise out.
Hello @crystalwizard, welcome to AI Forum!
Thank you for this feedback. You raise an excellent point regarding memory pressure. As Antigravity handles increasingly complex projects with large context windows, the local RAM footprint for indexing and agent orchestration can become significant.
I’ve flagged this for our team to review how we handle memory overflow in resource-constrained environments.
the other thing that would help a lot is a way to turn down the incredible push to ‘go fast and not think’ that all the gemini models are under. i have to constantly stop my agent and have it calm down, or even reboot the IDE. otherwise it gets into such a push to just CODE AT TOP SPEED! that it breaks everything it touches, it doesn’t even attempt to follow best practices, and then the entire project has to be scraped and restarted. go fast, optimize for speed - not good for small projects. really not good for large complicated projects that require multiple files and other sorts of things.