The individual inputs a model learns from
A feature is a single measurable input variable you feed into a model. If you're predicting a house's sale price, the square footage is a feature, the number of bedrooms is a feature, the age of the house is a feature. Each row in your dataset is one example, and each column, other than the thing you're trying to predict, is a feature describing that example. Nothing more mysterious than that.
But that simplicity hides the most important constraint in applied machine learning: a model can only find patterns in the features it's actually given. If square footage predicts price but you never included it, no algorithm, no matter how sophisticated, will magically discover it. The model isn't reasoning about houses, it's finding statistical relationships between the columns in front of it and the target. Practitioners sum this up as 'garbage in, garbage out.' A brilliant algorithm trained on weak or irrelevant features will lose to a simple algorithm trained on strong, well-chosen ones.