To add on to the image compression theory, I read somewhere (maybe here on HN) that perhaps many image memories are stored as difference from a "nominal" representation. So one might not remember a complete face for example, but rather the features that stood out from a "normal face".