You will need to store data which doesn't fit into RAM in secondary storage. But that doesn't mean you need a file system or even files.
> to be a useful abstraction
Files are a necessary evil in non-image based systems because you need to store data somewhere when the programs using them aren't running. As the different objects they contain, such as plain text, hypertext, photographs, sound recordings, executables, etc. have nothing in common, they seem an unnecessary abstraction. They require that programs which use them parse/serialize their contents. This is unnecessary if the contents are already in memory, already in the format the program needs.
> replicate the file-like conveniences of global tagging and sorting and organizing of all the objects in your image
Why remember a file name and where it is in the directory tree when you could use a search engine to search for it based on content? Or simply chain through objects, going to the field you want and following the link? Programs, of course, will just directly link to the object.
> work out the schemes for permissions
I'd go for capabilities, rather than access control.
> why not just teach it what a file is?
Building a file system is a major undertaking. If it can be avoided, and to the extent it can be avoided, it should be. You would only need to know about files when you interact with systems which are based around files.
Can you preserve this flexibility/loose-coupling feature in image-based storage? I don't know. I'd be interested to learn if it can be done so.
You could, with a non-trivial amount of effort, replicate the file-like conveniences of global tagging and sorting and organizing of all the objects in your image. You could, also with a non-trivial amount of effort, work out the schemes for permissions, etc. so that objects with all those file-like conveniences can be shared, like files, on multi-user systems or between machines or over networks.
It's not clear to me that the above would offer any tangible benefit over files. So if you're going to put a non-trivial amount of effort into a lisp machine, why not just teach it what a file is?