Sakhanda Wire
NVDA $153.77 -1.52% MSFT $504.91 -2.53% GOOGL $281.52 -0.78% META $620.99 -0.87% AMZN $256.80 -0.37%
← Back to the news

Stanford Researchers Release Memorilla Memory System for Frozen LLMs

Stanford researchers have released Memorilla, an open-source memory module designed to help frozen large language models use large document collections. The system compresses a collection into 16 learned memory tokens, which are then added to a question for Qwen3-8B. The decoder and Qwen3-Embedding-4B encoder remain frozen, while the 194,123,776-parameter memory module is trained.

On FactKG, the researchers report 84.33% accuracy with Memorilla, compared with 80.43% when all documents are placed in context, using 108 times fewer document tokens. Across eight reported benchmarks, it achieved an average score of 49.68 and led on five. Its strongest listed result was 44.60% on PersonaMem-v2, while PubMedQA was a weaker case at 59.00%, below LLMLingua-2’s 69.00%.

The public release includes code, training recipes and checkpoints under Apache 2.0. The system runs on Linux with CUDA and Python 3.11; the quickstart requires one 80GB GPU. Training the full recipe took about 19 hours on four H100 GPUs. The researchers also report that Memorilla answered queries in 0.74 seconds for 512 documents on one H100, compared with 19.9 seconds for full context.

Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.

Originally published by MarkTechPost on

Read the original on MarkTechPost ↗

The reporting and the image belong to MarkTechPost and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news