While parametric memory reveals what an LLM recalls unprompted from its internal training weights, grounded memory retrieves live web sources to show what the model extracts in real time, exposing gaps and distortions between static perception and current web reality.