▲ 62 today775 starsPythonHN ×1
Memorilla lets frozen LLMs remember whole docs with 16 tokens
What is memorilla?
Memorilla is a memory layer that compresses an entire document collection into a handful of query-dependent tokens for a frozen large language model. It feeds the model just 16 tokens per question, achieving higher accuracy than feeding all documents while using far fewer tokens.
Who it is for
Developers building chat assistants or retrieval-augmented generators can boost performance without expanding the model’s context window.
Try it in 60 seconds
You need: Linux with CUDA GPU and Python 3.11
git clone https://github.com/snap-stanford/memorilla.git && cd memorillauv venv --python 3.11 && source .venv/bin/activateuv pip install -e .uv pip install --no-build-isolation flash-attn==2.7.4.post1
Apache-2.0 licenseUpdated yesterdayPython5 open issues
New projects are not vetted. Skim the code, and run it in a fresh folder or a container before giving it your files or keys.
The story so far
Early Signal first spotted snap-stanford/memorilla on November 10, 2026, when it had 775 GitHub stars. It now has 775. On its busiest day it gained 62 stars. It is published under the Apache-2.0 license and has 62 forks. It has been discussed in 1 Hacker News thread, the top one with 6 points. Early Signal files it under LLMs, RAG & Memory, Coding.
Questions about memorilla
- Is memorilla free and open source?
- Yes. snap-stanford/memorilla is open source under the Apache-2.0 license, and the code is on GitHub at https://github.com/snap-stanford/memorilla.
- How do I install memorilla?
- Use one of these, depending on your setup: git clone https://github.com/snap-stanford/memorilla.git && cd memorilla · uv venv --python 3.11 && source .venv/bin/activate · uv pip install -e . · uv pip install --no-build-isolation flash-attn==2.7.4.post1
- What do I need to run memorilla?
- Linux with CUDA GPU and Python 3.11