Scope: Untargeted metabolomics may reveal preventive targets in cognitive aging, including within the food metabolome. Methods and results: A case‐control study nested in the prospective Three‐City study included participants aged ≥65 years and initially free of dementia. We contrasted 209 cases of cognitive decline and 209 controls (matched for age, gender and educational level) with slower cognitive decline over up to 12 years. Using a bootstrap‐enhanced LASSO regression on the baseline serum metabolomes analyzed with LC‐QTOF, we identified a signature of 22 metabolites associated with subsequent cognitive decline. The signature included three coffee metabolites, a biomarker of citrus intake, a cocoa metabolite, two metabolites putatively derived from fish and wine, three medium‐chain acylcarnitines, glycodeoxycholic acid, lysoPC(18:3), trimethyllysine, glucose, cortisol, creatinine and arginine. Adding the 22 metabolites to a reference predictive model for cognitive decline (conditioned on age, gender and education and including ApoE‐ε4, diabetes, BMI and number of medications) substantially increased the predictive performance: cross‐validated Area Under the Receiver Operating Curve = 75% [95% CI 70–80%] compared to 62% [95% CI 56–67%]. Conclusions: Our untargeted metabolomics study supports a protective role of specific foods (e.g., coffee, cocoa, fish) and various alterations in the endogenous metabolism responsive to diet in cognitive aging.