Some implementations of mipmapping store a single medium-resolution image and either sample fewer pixels or duplicate neighboring pixels when drawing. You don't need extra memory for that. Literally just mapping a subset of the pixels from the texture stored in memory.
When enough pixel duplication occurs you may want to enqueue a request for a higher resolution texture to be loaded into memory and let the resource pool decide. You can see this in effect in some older games where you're able to "outrun the horizon/frustrum" so to speak or when too many objects spawn and the level of detail is terrible for a while until the game catches up. Otherwise this is just a most-recently-used cache. Mipmapped textures can either be many prescaled static copies on disk or precomputed from the single full resolution asset on disk while loading a level.
When enough pixel duplication occurs you may want to enqueue a request for a higher resolution texture to be loaded into memory and let the resource pool decide. You can see this in effect in some older games where you're able to "outrun the horizon/frustrum" so to speak or when too many objects spawn and the level of detail is terrible for a while until the game catches up. Otherwise this is just a most-recently-used cache. Mipmapped textures can either be many prescaled static copies on disk or precomputed from the single full resolution asset on disk while loading a level.