The persistent spread of a destination host is the number of distinct sources that have contacted it persistently in predefined t measurement periods. A persistent spread estimator is a software/hardware component on a router that inspects the arrival packets and estimates the persistent spread of each destination. This is a new primitive for network measurement that can be used to detect long-term stealthy malicious activities, which cannot be recognized by the traditional super spreader detectors that are designed only for "elephant" activities. However, the challenge is to function such an estimator in fast but small memory space (such as on-chip SRAM of line cards), in order to keep up with the high speed of switching fabric for packet forwarding. This paper presents an implementation that can use very tight memory space to deliver high estimation accuracy: Its memory expense is less than one bit per flow element in each time period, Its estimation accuracy is over 90% better than a continuous variant of Flajolet-Martin sketches, Its operating range to produce effective measurements is hundreds of times broader than the traditional bitmap. These advantages originate from a new data structure called multi-virtual bitmap, which is designed to estimate the cardinality of the intersection of an arbitrary number of sets. We have verified the effectiveness of our new estimator using the real network traffic traces from CAIDA.