Memorilla: Memory System Cuts Tokens 4000x, 27x Faster Than Full Context
ShirleyYXWu · x · 2026-11-06
Memorilla is a new LLM memory system that converts long textual inputs and agent trajectories into latent inputs for a frozen LLM, no training required. The authors claim up to 4000x fewer input tokens and 27x faster inference than full-context, while staying accurate — pitched as memory as a modality for LLMs.
Related event: Memorilla: 16 Memory Embeddings Replace Full-Context LLM Inference(2 posts)→
More from coding & agent
- Perplexity Computer builds 6-minute GTA RAGE engine explainer from 40 web sources — jeffgrimes9 · 2026-11-06
- Agents still claim tasks will take 'several hours' — then finish in under 10 minutes — jacalulu · 2026-11-06
- Alchemy teases next-week release: Rivet and Effect Cluster as swappable deploy layers — samgoodwin89 · 2026-11-06
- Grok Build v1.0.75 ships major reliability and performance upgrades for multi-agent workflows — elonmusk · 2026-11-06
- Gauge open-sources CLI as 85% of coding sessions are now agent-initiated — ycombinator · 2026-11-06
- kaigen3d: a C-based agent-native game engine with headless operation and automated QA — gdechichi · 2026-11-06