One thing that has been constantly making tests hard for me is the way Memrise determines/presents alternative answers that are shown in multiple choice tests.
The system seems to preferably find/show other possible (wrong) answers that are similar to the “right” answer.
To find these, Memrise probably uses some fuzzy search algorithm (soundex, DMP, Levenshtein, etc.) to order entries by similarity and then take the first 4-8 words and present them.
Also, Memrise does not consider the type of a word or phrase - it presents a random mixture of i. e. nouns, adjectives, even sentences, etc.
IMHO it would be much better to present items of equal (or similar, FWIW) type, at least where available (most larger courses have a column that specifies the type). IOW, Memrise should allow course creators to instruct Memrise to only pick answers from other database entries with the matching type.
Both above factors result in tests becoming very hard in large courses where authors need to add info to items in order to distinguish them from another, with items/answers like the following:
"your [not thy, thine, yours, of you]"
(Actually I would suggest that there is a “not this and that” column that would allow for a comma-separated list of such words.)
When each of the above words also has their own entry, along with all others on the “not”-list, guess what happens: yes, you’ll (most probably) be presented with exactly those words in MPC tests (at least on Android which is what I’m using most of the time).
Here’s a course where this is making it next to impossible to work through speed reviews:
(course forum: [Course Forum] 5000 most frequent Dutch words ♫ Audio)