I've been watching your memcached process for a little while now and it's almost like it's leaking memory because it's closing in on 30 MB. I'm not sure what's causing your problem but given the following description upgrading all of our servers to memcached 1.2.2 couldn't hurt:

2007-05-02: Version 1.2.2 released. Memcached can be configured to support multithreading. Big code cleanup with performance improvements. Memcached now collects eviction and per-object-type statistics.

A memory leak was fixed in version 1.2.1 but Danga didn't mention it in their mailing list or website announcements, hopefully the same has happened for version 1.2.2.

Remi has upgraded memcached on all servers to version 1.2.2.

I realize this thread has been dead over over a year now, but I must say that I've also run into this problem.  I've set memcached to run with a limit of 10 mb, yet it creeps up over that amount within a day or two and eventually must push me over my limit for shared hosting 1 because it mysteriously disappears at some point

I'm only running a blog, powered by Django, so it's no huge deal (although I'm using memcached as a session backend, so while the blog is still viewable when memcached is down, no one can sign in.)  However, it would be great to find out how to keep memcached within its limits, after all, if -m is ignored then why even bother with that option in the first place?  Perhaps memcached isn't meant for such small caches.  What little I was able to find on the memcached mailing lists seemed to suggest that memcached will go over the limit for non-cache related memory (e.g. connections.) In that same response I also found:

There's one other place where memcached can exceed its memory limit: if a "set" request comes in whose object size requires a fixed chunk size that hasn't been seen yet, memcached will always allocate a 1MB slab for objects of that size, even if the cache has already reached its size limit. That ensures that you don't get bogus "out of memory" errors after filling up your cache with, say, 1000-byte objects, then sending in a 1500-byte object.

From memcached overrides memory limit

I'm not exactly sure what "fixed chunk size" refers to... surely memcached doesn't allocate 1mb to every value of a size that it hasn't encountered yet?

Anyway, I'll try fiddling with cache timeouts and running memcached with the verbose setting to see what's happening before I resort to a periodic CRON job to kill and restart memcached.

I've had this issue as well, so I thought I'd post what I've found.

I'm not exactly sure what "fixed chunk size" refers to... surely memcached doesn't allocate 1mb to every value of a size that it hasn't encountered yet?

This is actually what happens - not an issue when you've got a 2GB cache, bit more of a problem when you're running in Shared1 and trying to stay < 10MB.

This script will show you details of what slabs have been allocated and what's in them. Mine used to look like this:

 #  Item_Size   Max_age  1MB_pages Count   Full?
  2      60 B    51771 s       1     633      no
  3      76 B    51771 s       1    3739      no
  4      96 B    51771 s       1     309      no
  5     120 B      101 s       1       1      no
  8     240 B    40606 s       1       2      no
  9     300 B     5043 s       1       2      no
 10     376 B    38892 s       1       1      no
 11     472 B       62 s       1       1      no
 12     592 B      101 s       1       1      no
 13     740 B      456 s       1       2      no
 14     928 B      452 s       1       2      no
 15     1.1 kB     101 s       1       1      no
 16     1.4 kB     101 s       1       1      no
 20     3.5 kB   50734 s       1       1      no
 21     4.3 kB   38883 s       1       4      no
 24     8.5 kB   49086 s       1       1      no
 25    10.6 kB   49090 s       1       2      no
 36   123.4 kB     556 s       1       1      no

I had set m=10 but here you can see memcached has allocated 18MB of RAM, but most of the 1MB chunks only have 1 key in them (there's only about 500KB data in there in total). So, I bumped up the growth factor for the item sizes from the default of 1.25 to 8. This means that, for the larger items you end up wasting more memory per item, but because I've got so few items, this is nothing compared to the memory savings from getting rid of all those near-empty slabs.

[bios@web38 bin]$ memcached -l 127.0.0.1 -m 10 -p 2553 -n 64 -f 8 -d
[bios@web38 bin]$ ./memcached-tool 127.0.0.1:2553 display
  #  Item_Size   Max_age  1MB_pages Count   Full?
  1      96 B      139 s       1     162      no
  2     768 B      109 s       1       6      no
  3     6.0 kB     139 s       1       5      no
  4    48.0 kB      27 s       1       2      no

Which is waaaay better. You'll want to tweak f and n depending on the distribution of the sizes of the items in your cache.