Lunr.js [1] used to use a trie data structure and memory usage was a concern. The problem is that although the data structure compresses prefixes, it does nothing for suffixes. Depending on the corpus this can lead to a lot of duplication.
Switching to a trie-like structure that compresses prefixes and suffixes can lead to significant savings. Building the structure can be a bit more burdensome, so there is a trade off there. There is a paper describing the approach [2] and, if you're interested, my JavaScript implementation [3][4].