Memory utilization is how much of a machine's physical RAM is currently committed, normally reported as used bytes, available bytes, total bytes, and a percentage derived from them.
On Linux the used figure on its own is close to useless, because the kernel deliberately fills spare RAM with page cache: cached file data that is instantly reclaimable the moment a process needs memory. A perfectly healthy server can sit near 95% used forever and never be short of anything. The number to watch and to alert on is available (MemAvailable), the kernel's own estimate of how much can be handed to a new allocation without reclaiming aggressively. Available memory sliding down over hours is the shape of a leak; a single dip is usually just a large file being read.
Memory problems end abruptly rather than gradually. The OOM killer terminates the biggest offender and the process disappears with no slow ramp in response time first, which is why an external check often catches it only after users do. Hyperping records used, available, and total memory with utilization percentage on a 30 second cadence, so the downward slope of available memory shows up well before the kill.
