> So at inference time, it does not “learn” in the training sense

It learns because it remembers the context. The larger the context, the better the capabilities of the model are. I mean just give it a try and see for yourself - start building a feature, then next feature, then the next one etc. Do it in the same "workspace" or "session" and after few days, one or two weeks of writing code with the agent, you will notice that it somehow magically remembers the stuff and builds upon that context. It becomes slower too.

"Re-learning" is something different and it may not be even needed.