Looking to add_kubernetes_metadata, it also has a cache, that relies on receiving delete events from the Kubernetes API to delete its entries. The problem could be there, there are cases where delete events are not always received. These caches would stay in memory forever.
Update: Pods are also removed from the cache if they are marked for termination in an update. In any case it'd be good to check if the memory issue is there.
@mareckii it'd be great if you could confirm in your scenario if the problem is with add_process_metadata or with add_kubernetes_metadata.