Such “stickyness” is a property of a model as a whole, not any individual parameter in isolation. Hard to say much without seeing the full model.
I will say this: lognormal(0.0, 1e6) is a preposterous distribution. Remember that that sigma is on the log scale, beta1 could easily be as large as e^{10^6} or as small as e^{-10^6}!
Try lognormal(0.0, 10.0) if you need a diffuse prior. That still covers several orders of magnitude.
Thanks, nhuurre. I just realize that the 1e6 might be one reason. I try several numbers for lognormal. 10.0 still cause the problem. 1.0 seems ok but the parameter value is a little bit small. 5.0 seems ok.