snap-stanford/memorilla: Latent Semantic Memory for LLMs

Memorilla is a memory system designed for frozen large language models that compresses entire document collections into a fixed number of query-dependent memory tokens. Researchers report that Memorilla outperforms existing methods like Mem0 and MemoRAG while drastically reducing token usage and response times. The project code, checkpoints, and evaluation guides are available under an open-source license. Open article →