Apple's MemoryLLM adds interpretable memory to transformers

Apple ML Research introduces MemoryLLM, a plug-and-play interpretable feed-forward memory module for transformers. The work aims to improve interpretability of feed-forward networks, which are core to recent LLM advances.