"We have each accumulated enough room-frames to represent most rooms we're likely to see". But Minsky goes one step further than this, and proposes a generic frame that fits almost any room, with terminals corresponding to ceiling, floor, and walls. Then, each of these terminals is represented by a subframe that includes direction-nemes to identify different regions of the surface. The point of this is to use the direction-nemes to encode the spatial relations of objects that might be seen in different views.
But what about mistakes -- like misidentifying an object type? This kind of mistake won't be catastrophic if related frames share the same terminals. The frame representation can be easily revised, preserving the other correct representations that have been built.
NB: This approach is interesting, but it buys into a specific viewer-centered coordinate system. How might the agency approach be used to represent, say, the object-centered coordinate systems of the objects in the room? Would the same kind of direction-neme strategy work?