snap-stanford/memorilla

★801PythonApache-2.0⑂ 62Issue 2Updated Nov 10, 2026

View on GitHub ↗Website ↗

🦍 Memorilla: Latent Semantic Memory for LLMs

Compressing Entire Document Collections into 16 Memory Tokens — Memorilla is a Stanford research system that condenses an entire document collection into a fixed set of query-dependent memory tokens for a frozen LLM. With just 16 tokens, it beats Mem0 and MemoRAG on average and even outperforms feeding every document on FactKG (84.33% vs 80.43%) using 108x fewer tokens, while also boosting an agent's TextWorld success rate from 18.6% to 24.2%.

Metrics

Retention can be computed 30 days after a repository first trends (it isn't shown until then). See About for how each metric is defined.

Star growth and Trending appearances (last 30 days)

Created Jun 23, 2026 · First on trending Nov 10, 2026

Other notable repositories in the same category (LLMs & ML)

• kerlenton/bootgptA Tiny Neural Net That Boots Straight From the MBR
• scienthoon/jev-ood-calibrationTesting Whether Jev's Confidence Scores Stay Honest Out-of-Distribution
• alejandrohdez00/alien-space-of-scienceSampling research ideas that are coherent but hard for humans to reach
• microsoft/VibeVoiceMicrosoft's open voice AI grows into streaming speech recognition
• neuml/txtaiA Single Framework Combining Vector Search, Graphs and LLM Agents

See all LLMs & ML repositories →

Rank within llm

See all llm repositories →