We've had this conversation before, too. It's the object cache, and the process runs at a very low priority. Free memory is wasted memory. If you start thinking in terms of all volatile memory acting as cache, it makes sense: CPU caches (L1, L2, et al) <= RAM <= nonvoltile memory (disk). With a 64-bit virtual address space, it even makes sense to report nonvolatile storage as "memory" and ignore RAM completely when not talking about caching. This makes even more sense on systems like the iPhone, where the traditional file system model has been replaced by application-specefic datastores transparently accessed locally or via "the cloud." Eventually, even disk storage will simply be cache. As storage performance improves, the line between RAM and disk gets fuzzy anyway.
EDIT: Some perspective. The DRAM used in classic Amiga accelerators has a maximum throughput of about 200 MBps. The lowliest of typical SSDs has a maximum read throughput of about 170 MBps, while the best approach 250 MBps. Access times in SDDs are quickly improving, as are peripheral buses. It won't be long before nonvolatile storage is indistinguishable from local RAM in terms of performance. When that happens, RAM will either be displaced completely (you can reserve a portion of disk for the purpose), or it will be used as cache. Since storage devices already have cache, RAM could become completely redundant, depending on the needs of the system. Some implementers may prefer to keep caching algorithms under the control of the operating system.
