Stanford SNAP group (Jure Leskovec), with co-authors from Toronto, Penn and MIT · Latent Memory Module

Memorilla

Memorilla is a 194M-parameter latent memory module that encodes a document collection into 16 memory tokens for a frozen Qwen3-8B decoder, with a frozen Qwen3-Embedding-4B encoder. On FactKG it scored 84.33% against 80.43% with full context, using 108x fewer document tokens, and it reports the highest average (49.68) across eight benchmarks, though it trails MemoRAG on some tasks and PubMedQA. The benchmark scores are self-reported. Code, training recipes and checkpoints are released under Apache 2.0, and the team also released the PersonalizationV4 dataset under CC BY 4.0.

Open the source ↗

Similar Research & Safety releases

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.