There is a management API which might provide this kind of information.
 Be warned though that it is not intended for usage from within the JVM
and likely causes significant overhead.

Why do you "need" memory usage?

    robert

Hmm, maybe instrumentation could help?

Simple example on how to compute approximated object size is there:
http://tinyurl.com/hfzzu
(notice the correction from a post next to it)

Accuracy of that method is much better than with serialization, and
efficiency is not so bad I think.

For example, on my machine ArrayList filled with 1000000 new Objects
consumes about 14MB of memory, and computation of its size takes about
1.5 sec.

piotr

It's good idea but cause large overhead :/

My recent experience with serialization is that what was 700 bytes
serialized using the default implementation took only 50 bytes or so
after I wrote my own custom serialization for the same objects. (This
matters to me because I want to fit it into one single UDP packet and
about 500 bytes seems to be the suggested maximum payload.)

What I left out, I expect, was administrative data for the objects
contained within the topmost object (and most of the administrative
data for the topmost object itself). My custom method would just write
the floats, ints, etc., contained in each nested object to an
ObjectOutputStream rather than write the actual objects themselves.

I expect my custom serialization is a more accurate estimate of the
in-memory cost of these objects than what the default serialization
was. The accuracy of the approach you suggest will therefore depend
entirely upon how the data objects in question are structured: if
there is much administrative data to be serialized, your estimate
could easily be off by a factor 10.

Cheers
    Bent D