Clinical diagnostic reasoning is a sequential process: a working hypothesis is formed early in an encounter and updated as new evidence arrives, and biases at one stage can compound across the stages that follow. In human clinicians, anchoring and premature closure account for a large share of preventable diagnostic errors. As large language models are increasingly considered for clinical decision support, an analogous question arises: how do they handle wrong context that enters mid-encounter - whether from their own prior outputs, an earlier clinician’s note, or pushback from a user? Existing evaluations of LLM cognitive bias use single-shot question answering and cannot speak to this dynamic. We introduce a temporal-replay benchmark over MIMIC-IV in which 686 patient timelines are replayed step by step to three open-weight LLMs, with each step’s structured assessment carried forward as the prior for the next. At three clinically meaningful anchor points we inject the same wrong diagnosis through four structurally distinct channels. The most damaging is corruption of the model’s own previous assessment, with effects roughly twice those of any other channel. When this manipulation derails the correct answer, the model lands specifically on the planted diagnosis about half of the time.