MemoryLLM: Plug-n-Play Interpretable Feed-Forward Memory for Transformers

Understanding how transformer components operate in LLMs is important, as it is at the core of recent technological advances in artificial intelligence. In this work, we revisit the challenges associated with interpretability of feed-forward modules (FFNs) and propose MemoryLLM, which aims to decouple FFNs from self-attention and enables us to study the decoupled FFNs as context-free token-wise neural retrieval memory. In detail, we investigate how input tokens access memory locations within FFN parameters and the importance of FFN memory across different downstream tasks. MemoryLLM achieves context-free FFNs by training them in isolation from self-attention directly using the token embeddings. This approach allows FFNs to be pre-computed as token-wise lookups (ToLs), enabling on-demand transfer between VRAM and storage, additionally enhancing inference efficiency. We also introduce Flex-MemoryLLM, positioning it between a conventional transformer design and MemoryLLM. This architecture bridges the performance gap caused by training FFNs with context-free token-wise embeddings.

Feed-forward components of LLMs, which hold roughly two-thirds of all model parameters, operate on a complex mixture of signals from other attention and residual modules of network, making it nearly impossible to cleanly study what they're doing. We designed MemoryLLM, a new architecture that completely isolate feed-forward modules from the rest of the network and trains them exclusively on individual word embeddings. Tis design improves interpretability and provide practical paths toward more transparent, resource-efficient AI systems which can be inspected, edited, or compressed in a principled way.