Indeed... Lindens are (for most of them) far from being optimization champions, like most "young" (30-40 years old and younger) coders who never learned to truly optimize code: back in my first coding days (in the early 70s), the most powerful "computers" I coded on had 1Kb of RAM, 4Kb of ROM, and a 1MHz CPU... This *does* make you learn to squeeze *every byte* and *every CPU clock cycle* out of your code. By my standards, most of today's so-called programs are just bunches of inefficient, bloated crappy pieces of code... That's why you need 2GHz multi-core CPUs and 8Gb of RAM today to "run" (more like "crawl") modern OSes and software...
There were even worst things in the original code for the v1 viewers (that Lindens finally learned to take into account in their newer code), such as saved settings variables (AKA debug settings, which are in fact used for far many more purposes than just debugging), that are referenced via strings (meaning that to retrieve the value of a simple boolean from a setting, you must first compare the name of its variable, character per character, with the names of all the variables) and that only "recently" (IIRC, back in Snowglobe days) had their value cached (LLCachedControl) for faster access.
See by yourself... Compare the memory usage of v1.26.8.29 with v1.26.8.30, for example (in the exact same conditions, of course, i.e. same settings, same UI elements on screen (e.g. same open floaters), same 3D environment, preferably in an empty sim to avoid issues with nearby avatars coming in and out between the two sessions).
Your mileage may vary also depending on the memory allocator in use (the figure I gave was for Linux with jemalloc as the memory allocator): the gain will be different if the memory allocator reserves space on a byte-aligned scheme, like under 32bits Linux without jemalloc or under Windows without tcmalloc (in which case each C++ object will use exactly the amount of memory needed by its member variables) when compared to 16-bytes aligned allocations (jemalloc, tcmalloc, MacOS-X, where the space allocated to each object is rounded up to the nearest 16 multiplier).
Yes, I already explained that countless times on this forum...