Bundle
Adding several vectors into one of the same size that still resembles all of them.
Like you're twelve
Play three notes at once. What comes out of the piano is one sound, and a trained ear can still pick out each note inside it. The chord is not the average of the notes — it is all of them, at the same time, in one signal.
Bundling is the chord. Three ideas go in, one vector comes out, and the vector is still recognisably each of the three.
Precisely
Bundle is weighted summation, normalised by default. The result has non-trivial similarity to every input, which is the exact opposite of what binding does — and the two together are what make records representable: bind each role to its filler, bundle the bindings, and you have one vector for the whole record.
Weights may be negative, which is subtraction rather than an error: bundling with a negative weight removes a concept from a mixture.
Capacity is finite. Each additional term adds interference, so a bundle of a handful of items reads back cleanly and a bundle of hundreds does not. This is Superposition, and the ceiling on it is set by dimensionality.