Yes, you can create an alarm for any single metric. In this case, the single metric has a dimension that represents all instances. Here's how you can do it in boto.
In [1]: import boto
In [2]: cw = boto.connect_cloudwatch()
In [3]: cw.list_metrics(metric_name='CPUUtilization')
Out[3]:
[Metric:CPUUtilization,
Metric:CPUUtilization,
Metric:CPUUtilization,
Metric:CPUUtilization]
In [4]: l = _
In [5]: for m in l:
...: print m.name, m.dimensions
...:
CPUUtilization {u'ImageId': [u'ami-1b814f72']}
CPUUtilization {u'InstanceId': [u'i-366c4354']}
CPUUtilization {}
CPUUtilization {u'InstanceType': [u'm1.large']}
You can see here that there are four separate metrics associated with the metric_name CPUUtilization. The first has a dimension of all instances that use that particular AMI, the second has a dimension for a particular instance that is now running, the fourth has a dimension of all instances of a particular type, but the third has no specified dimension. This metric represents CPUUtilization across all of my instances. So:
In [6]: m = l[2]
In [7]: m.create_alarm(name='cpu_all_instances', comparison='>', threshold=80.0, period=60, evaluation_periods=2, statistic='Average')
Out[7]: MetricAlarm:cpu_all_instances[CPUUtilization(Average) GreaterThanThreshold 80.0]
This alarm should fire if the average CPU utilization across all my instances exceeds 80% for two evaluation periods. You could also choose a different statistic, like 'Maximum', that would fire if the maximum value for CPU utilization across all instances exceeded 80% for more than 2 evaluation periods.
I don't know if you are specifically looking for MemoryUsage or if that was just an example but I don't think MemoryUsage is one of the available metrics from CloudWatch.