Can a "frozen" model follow rules from external memory better than those from its own context window?

Can a frozen LLM follow complex rules stored entirely outside its context window better than when reading those same rules directly in its prompt? In our experiments with a completely frozen Phi-2 (2.7B), the model achieves 98.4% legal moves (179/182) in a card game when rules, examples, and correction notes exist only in external latent memory—dropping to 0–1.6% when that memory is detached, and outperforming the exact same model reading the material directly in its context window (98.4% vs. 88%). On held-out fact retrieval outside the context window, it scores 47/60 (vs. 29/60 in-context and 0/60 without memory).