Stanford Researchers Release Memorilla, an Open-Source Memory Module for Large Language Models
A team led by Stanford's SNAP group has introduced Memorilla, an open-source memory system that condenses entire document collections into just 16 tokens for frozen language models.
Key points
- Stanford researchers released Memorilla, an open-source memory system for frozen LLMs.
- The system condenses document collections into a fixed output of 16 memory tokens.
- It achieved strong benchmark scores while using significantly fewer document tokens than full-context methods.
Researchers at Stanford University, alongside co-authors from Toronto, Pennsylvania, and MIT, have released Memorilla, a new memory architecture designed to expand the information handling capabilities of large language models without altering the base network. Developed by Jure Leskovec’s SNAP group, the system processes massive document collections and outputs a fixed set of 16 memory vectors that are then fed into a frozen language model. The project's code, training recipes, and model checkpoints have been published under the Apache 2.0 license.
Traditional approaches to handling large text corpora either strain context windows or rely on retrieval systems that can miss subtle clues across disparate files. Memorilla takes a different approach by utilizing a trainable 194-million-parameter memory module. The pipeline features a frozen Qwen3-Embedding-4B encoder to process documents once, a reading phase where learned slots attend to the entire document store, and an injection step that prepends the resulting vectors into the input space of a frozen Qwen3-8B decoder.
According to the research team, this setup achieved an 84.33% accuracy score on the FactKG benchmark using 108 times fewer document tokens than standard full-context prompting, which scored 80.43%. Across eight evaluated benchmarks, the system achieved the highest reported average score of 49.68, outperforming alternative methods like xRAG and Mem0, though it lagged behind specialized tools like LLMLingua-2 on medical reasoning datasets such as PubMedQA.
What to watch next: Look for adoption of Memorilla in enterprise search environments and further testing of latent memory modules on larger commercial language models.
Source MarkTechPost — marktechpost.com/2026/11/14/stanford-memorilla-memory-system-16-mem… ↗
Read next
🚀Startups & FundingHarvard Study Reveals US Hyperscale Data Centers Consume Power Rivaling Entire States
A new Harvard study reveals that America's 403 hyperscale data centers consumed between 68 and 99 terawatt-hours of electricity during the 2024-25 study period, rivaling …
🚀Startups & FundingCalifornia Files Lawsuit Against OpenAI Over Autonomous AI Incidents
California Attorney General Rob Bonta has filed a lawsuit against OpenAI, alleging the company failed to control its autonomous AI agents and misled the public about safe…
🧠AI AnalysisPedestal AI Secures Seed Funding to Automate Logistics With Intelligent Agents
Pedestal AI has launched publicly with $4.7 million in seed backing to deploy agentic artificial intelligence aimed at streamlining logistics and supply chain operations.
🏢Big TechUnited Launch Alliance Investigates Vulcan Rocket Failure
United Launch Alliance is investigating a mission failure of its Vulcan rocket, which suffered from low booster performance shortly after liftoff. The incident raises que…
🌐Work & SocietyUS Southwest Hydropower Output Hits Record Lows Amid Persistent Drought
Hydropower generation in the US Southwest dropped 28% below the 10-year average this summer as record-low water levels in the Colorado River Basin threaten regional energ…
🏢Big TechYouTube Updates View Metrics to Include TV Co-Viewing Estimates
YouTube is incorporating estimated co-viewing data into its public view counts to better reflect group consumption on television sets.