We integrate dispersed information into a heterogeneous-agent New Keynesian model in which households cannot fully distinguish idiosyncratic from aggregate income, creating a two-way feedback between household sentiment about future income and aggregate demand. We develop a portable methodology—based on a forecast matrix—to embed various types of learning into arbitrary sequence-space models. Our model yields three key results: delayed overreaction of beliefs to shocks, a “sentiment inertia” feedback that amplifies hump-shaped responses, and a “sentiment accelerator” that generates persistently slow recoveries. These results depend crucially on the interaction between household heterogeneity and dispersed information.