I definitely think that it's good for programmers to understand memory management and how the system actually works... but, I think ARC is a very good system and works really well. This is really an opinion question, so my opinion is that it's almost always worth starting new projects that are going to target iOS 5 apps on ARC, except in very specific circumstances.
I feel that if you're using a lot of C libraries in your code, ARC is a little bit harder to use right now (so if you're mostly using third-party C libraries and things like CoreFoundation, you might consider whether it makes sense or not), but even then, if these libraries are mostly isolated from your Objective-C controllers and such, ARC is still good.
For older apps, you need to look at your app usage and patterns. If you use a lot of delegate methods, since you can't use weak references on iOS 4, it gets a little more tricky and you'll probably have to have mixed ARC and non-ARC code. It might be better to make a design decision to move forward with ARC. So new features are designed for iOS 5 and maybe not available (or fully available) in the iOS 4 version of the app, and those use ARC.
Really, in the end, it will depend on how your application is already designed, how large it is, and how comfortable you are with managed memory management and the usage/restrictions of ARC. For example, I have three projects that I'd never convert to ARC, one that I'm doing mixed right now, one that's fully converted (but still targets iOS 4+) and 2 that are full-on ARC and iOS 5+ only.