Researchers added a per-token integer fact-table to frozen GPT‑2, integrating it into the model’s logits via confidence weighting to provide live memory beyond the native 1,024-token window. With facts inserted roughly 1,500 tokens back, GPT‑2 alone scored 0/8 on both recall and revision, while the augmented model achieved perfect 8/8 scores on both tasks. This demonstrates that a lightweight memory organ can extend a transformer’s effective context without modifying