Probably, the only general rule to follow is that the key should be something that well characterizes the document. Anything more specific will depend a lot on the topics you are writing about, and vary from reference to reference:
- Is the title very unique, or rather generic? In the former case, a shortened version of the title could be integrated into the key. (As a random example: The title The triangle processor and normal vector shader: a VLSI system for high performance graphics might be shortened to something like TriangleProcHighPerf.)
- Does the publication describe a product or technique that has a name (which you might also use in your text)? If so, that name could become a part of the key.
- Is the publication connected to a recognizeable author name, or do you rarely see the same author name twice in the literature you deal with, in particular with respect to specific approaches? In the latter case, author names may just be arbitrary strings that do not help you remember anything particular, while in the former case, you might think about including the author name the work is associated with the name in the key.
- Is the year in any way special for the work? For example, is it an exceptionally early example of a supposedly modern invention, or is it the variant that has become known as "the 2011-version" of a particular approach? If so, the year could reasonably be a part of the key, otherwise, it seems superfluous.
- Can the publication be categorized? For instance, you may want to indicate in the key whether something is a concept draft, a user study of a concept presented elsewhere, a survey of several techniques, or a design rationale for a given concept.
- Are there various versions of essentially the same work published by different publishers? Different layouts and presentation forms (monochrome vs. color, ...) may have different strengths, so you may end up wanting to specifically refer to (w.l.o.g.) the Springer version and the IEEE version of some work that for some reason was published twice. In that case, including the publisher name in the key might be reasonable.
I hope this helps to get the idea - I do not see a reason for a uniform key format here; instead, this case-specific format highlights the peculiarities of each referenced work and therefore seems to help best to remember which reference points to what work in my experience.
For me personally, the above system generally leads to keys that never include an author name or publication year, and almost always a concept name, otherwise some fragments of the paper title. Depending on your topics, you may well end up with different preferences.
I have never had a problem of key collisions while using this approach; if anything (not that it would actually cause any problem), I may have ended up with several different keys in cases where I created bibliography entries for the same publication several times rather than copying the first one to later works.
