Were those 18-20 BMI people who you mention, the ones who always were
thin, or the ones who were fat and then became thin? I suspect that
these long lived people are the naturally thin ones.

Can a fat (not "naturally thin") person improve life expectancy by
losing weight in some sensible way? That's not 100% clear to me,
although there are good hopes for that (diabetes prevention study
etc). Anyhow, even if weight loss does not increase my life
expectancy, I would rather be a thin person than a fat person.

yep. family history and such.

A very large scale study (
http://content.nejm.org/cgi/content/full/341/15/1097 ) found that:

"the nadir of the curve for body-mass index and mortality was found at a
body-mass index of 23.5 to 24.9 in men and 22.0 to 23.4 in women."

The same study also noted that:

"The curve for the risk of death from cardiovascular disease was J-shaped;
for the risk of death from all other causes, the curve was U-shaped. The
J-shaped and U-shaped curves were explained primarily by an increased risk
of death among lean men and women as a result of cerebrovascular disease,
pneumonia, and diseases of the central nervous system (data not shown). A
high body-mass index was most predictive of death from cardiovascular
disease, especially in men (relative risk, 2.90; 95 percent confidence
interval, 2.37 to 3.56). Significantly increased risks of death from
cardiovascular disease were found at all body-mass indexes of more than 25.0
in women and 26.5 in men."

With respect to "lean" individuals the study noted:

"Limiting the primary analyses to subjects who had never smoked and who had
no history of disease at enrollment greatly reduced the apparent elevation
in the risk of death among lean persons, increased the risk among heavy
persons, and shifted downward the body-mass index level associated with the
lowest risk of death."

and,

"Thus, the increased risk of death from specific causes that was associated
with leanness in this and other studies may reflect preexisting, but
unrecognized, disease processes, even after careful exclusions have been
made. Also, although we had information on recent weight loss, we were
unable to control for long-term weight loss. Several investigators have
suggested that there is an association between illness-related weight loss
over a period of many years and the subsequent risk of death in very thin
persons and that controlling for recent weight loss may be insufficient."

Bottom line...thinner is generally better, with respect to mortality.  But,
answer to the question "How low should I go?" is not entirely clear.