Memorilla: Latent Semantic Memory for LLMs
The gist
Stanford researchers' Memorilla paper compresses a document collection into 16 query-dependent memory tokens for a frozen LLM.
It beats Mem0 and MemoRAG on five of eight benchmarks, cutting token use sharply.
How it unfolded
- Oct 15, 2026The paper "Memorilla: Latent Semantic Memory for LLMs" is posted on arXiv (arXiv:2610.21375).
- Nov 3, 2026The Stanford Data-Centric AI Workshop schedule lists Vignesh Kothapalli presenting Memorilla at 11:00–11:20.
- Nov 6, 2026Hacker News posts link to the arXiv paper and its GitHub page, and a summary item is published on proomt.
Key terms
- LLMs
- A machine learning model trained on huge amounts of text to generate, summarize, translate, and analyze language.
- embeddings
- A numeric vector representation of a word or phrase that encodes its meaning so similar items are close in vector space.
Source: Wikipedia