There are a couple of models for self-directed learning resources that are useful to look at, because historically they’ve been extremely effective.
One is the reference-material model. Its extreme end is a system like wikipedia: few things are more self-directed than a huge set of extremely detailed articles all hyperlinked together. If you have a goal in mind, you can search and probably find what you’re looking for if you know the right keywords; if you have no goal, you can idly wiki-walk and sate your curiosity indefinitely. (For media, tvtropes may be better than wikipedia, but tvtropes also has elements of the next model I’m discussing — the collegiate model — and may fit better into that category for various reasons. However, for very particular domains there are better resources than wikipedia: memory alpha, for instance, has better coverage of star trek trivia than wikipedia does.)
Another is the collegiate model. I call it that because this is the kind of thing that universities try to produce in both students and research faculty, and it’s also one of the arguments for think-taks and open-plan offices. I would argue that the most extreme form of it is in collaborative open question-answering communities like Quora and Stack Overflow, with comment-centric news sites with voting mechanisms (reddit, hacker news) and any other discussion system with large numbers of extremely active members (imageboards like 4chan for instance) coming in second. The point of a collegiate model is that you have people from different fields of interest interacting with each other in a dynamic way. Everybody learns, everybody takes the role of both teacher and student, and often the community itself produces new ideas and names them. A reference-material model is extremely effective at allowing a motivated learner to access existing well-documented ideas, but a collegiate model produces new ideas and allows motivated learners (even if they are beginners) to participate in the process of producing those ideas. That said, the reference-material model imposes some rigor on ideas (mostly by transplanting the rigor of existing established systems like peer review, academic publishing, canonicity, and journalistic standards), while the collegiate model’s flaws and benefits come from its flexibility. A collegiate model system will produce many ideas, and only a few of them will be good; furthermore, any bias in membership is likely to perpetuate itself (male-heavy communities often over time become misogynistic and then become even more male-heavy as they shut out women; the same is true of sample biases in terms of economic situation, race, religion, and even otherwise innocuous differences like tendencies toward logical positivism versus epistemic agnosticism or belief in free will versus determinism. For extreme examples, look at 4chan’s various boards, many of which are incredibly creative and have had a huge impact on the culture of the internet but each of which have consistent and absurdly biased cultures and habits that often don’t translate well between boards).
Attempts to automate self-directed learning have historically relied upon sticking a layer of gamification on top of the reference-material model — everything from hand-held trivia games to twenty questions to duolinguo do this (and memrise does this while sticking a tiny bit of collegiate model in a bag on the side by allowing people to create and vote on mnemonic devices).