We present a study of the redshift evolution of the projected correlation function of 593 X-ray selected active galactic nuclei (AGNs) with I AB < 23 and spectroscopic redshifts z < 4, extracted from the 0.5-2 keV X-ray mosaic of the 2.13 deg2 XMM- Cosmic Evolution Survey (COSMOS). We introduce a method to estimate the average bias of the AGN sample and the mass of AGN hosting halos, solving the sample variance using the halo model and taking into account the growth of the structure over time. We find evidence of a redshift evolution of the bias factor for the total population of XMM-COSMOS AGNs from to with an average mass of the hosting dark matter (DM) halos log M 0(h --1 M ) ~ 13.12 ? 0.12 that remains constant at all z < 2. Splitting our sample into broad optical line AGNs (BL), AGNs without broad optical lines (NL), and X-ray unobscured and obscured AGNs, we observe an increase of the bias with redshift in the range and which corresponds to a constant halo mass of log M 0(h --1 M ) ~ 13.28 ? 0.07 and log M 0(h --1 M ) ~ 13.00 ? 0.06 for BL/X-ray unobscured AGNs and NL/X-ray obscured AGNs, respectively. The theoretical models, which assume a quasar phase triggered by major mergers, cannot reproduce the high bias factors and DM halo masses found for X-ray selected BL AGNs with L BOL ~ 2 X 1045 erg s--1. Our work extends up to z ~ 2.2 the z 1 statement that, for moderate-luminosity X-ray selected BL AGNs, the contribution from major mergers is outnumbered by other processes, possibly secular ones such as tidal disruptions or disk instabilities.