We study halo assembly bias for cluster-sized halos .
Previous work has found little evidence for correlations between large-scale bias and halo mass assembly history for simulated cluster-sized halos , in contrast to the significant correlation found between bias and concentration for halos of this mass .
This difference in behavior is surprising , given that both concentration and assembly history are closely related to the same properties of the linear-density peaks that collapse to form halos .
Using publicly available simulations , we show that significant assembly bias is indeed found in the most massive halos with M \sim 10 ^ { 15 } M _ { \odot } , using essentially any definition of halo age .
For lower halo masses M \sim 10 ^ { 14 } M _ { \odot } , no correlation is found between bias and the commonly used age indicator a _ { 0.5 } , the half-mass time .
We show that this is a mere accident , and that significant assembly bias exists for other definitions of halo age , including those based on the time when the halo progenitor acquires some fraction f of the ultimate mass at z = 0 .
For halos with M _ { vir } \sim 10 ^ { 14 } M _ { \odot } , the sense of assembly bias changes sign at f = 0.5 .
We explore the origin of this behavior , and argue that it arises because standard definitions of halo mass in halo finders do not correspond to the collapsed , virialized mass that appears in the spherical collapse model used to predict large-scale clustering .
Because bias depends strongly on halo mass , these errors in mass definition can masquerade as or even obscure the assembly bias that is physically present .
More physically motivated halo definitions using splashback should be free of this particular defect of standard halo finders .