Researchers have proposed a method for transferring knowledge between large language models using an external memory called Engram. This memory stores learned information in a table that can be accessed through a small reader. The study investigates what matters more when moving this memory to a different model: the frozen memory itself or the target-side reader. The results show that both the memory content and correct addressing matter, but the transferred table only becomes useful with a reader aligned to the target model. A lightweight reader can nearly close the gap in performance between using the same model and transferring knowledge across models. --- Why it matters: This research is important for engineers working on large language models because it provides a way to transfer knowledge between different models, which could improve their performance on downstream tasks such as question answering. The results also suggest that target-side adaptation can further improve alignment when direct reader reuse is insufficient. Source: https://arxiv.org/abs/2608.17050