Hello
I am experimenting with the PyTorch C++ frontend & noticed some unexpected memory behavior when working with tensors. 
Specifically; when I create tensors inside a loop and pass them to functions; I sometimes see memory usage growing continuously, as if the tensors are not being released. 
This does not happen in my Python experiments; so I suspect I am missing something about how the C++ API handles reference counting and tensor ownership.
I tried using at::Tensor::detach() and also checked whether torch::NoGradGuard could help, but the issue persists in long-running processes. It is not always easy to reproduce in small code samples but in larger projects, the memory growth becomes noticeable after running many iterations. 
Checked C++ — PyTorch 2.8 documentation guide for reference. I am also curious if insights from PyTorch C++ memory handling could be applied when learning How To Build AI Agents efficiently.
I want to understand if this is a misuse of the API on my side / if there are specific cleanup steps required in C++ that are handled automatically in Python.
Has anyone else run into this? If so; what is the recommended pattern for managing tensor lifetimes safely in C++?
Any best practices / examples would be very helpful; since good memory management is crucial for production C++ projects using PyTorch.
Thank you !!