The condition you describe fits the SQLOS memory manager bug that is fixed in CU4 (http://support.microsoft.com/kb/2845380) but if you are going to take an outage to apply CU4, I would just go ahead and apply CU6 and bring your instance up to the latest update. You would have to track your resource monitor ring buffer entries to get more information about the internal low notification being set that is trimming memory internally to troubleshoot further, but it would generally be easier to test the CU on a test server under your normal workload to see if the behavior changes to what is expected in most cases.