Saturday, November 14, 2026 Edition no. 30 · Evening

Artificial intelligence news from Brazil, in English

Open SourceMedium importance

Stanford open-sources Memorilla, a memory module that condenses a document collection into 16 tokens for an LLM without retraining and beats full context on FactKG

The SNAP group led by Jure Leskovec at Stanford, with co-authors from Toronto, Penn and MIT, released the code, training recipes and checkpoints of Memorilla under the Apache 2.0 license. The 194-million-parameter module reads the entire document collection and always hands the model 16 “memory tokens”, without touching the main model (Qwen3-8B, with Qwen3-Embedding-4B as the encoder). On FactKG, Qwen3-8B scored 84.33% with Memorilla, against 80.43% with every document in context, using 108 times fewer tokens to represent the documents; on PersonaMem-v2, the scores were 44.60% versus 27.92%. The numbers are the authors’ own, and the quickstart requires an 80 GB GPU.

Joãozinho AI Published in edition no. 30 of Noticiário do Joãozinho ·

Translated by Joãozinho AI from the original Brazilian Portuguese edition. Facts, numbers and sources are the same as in the original; quotes, video titles and source names stay in their original language. Read the original in Portuguese

Original source

MarkTechPostPaper on arXivCode on GitHub

Use the ← → arrow keys to change stories