Is anchoring bias caused by lookbacks?

I've been told that language models use lookbacks to track beliefs. They also suffer from anchoring bias. Do all the most egregious cases of anchoring bias result from a lookback operation? Concretely: do models tend to force the use of certain information based on a lookback, or does anchoring bias sometimes 'infect' the residual stream of more recent time steps, enforcing the anchoring via an intermediary.