Stanford's Memorilla: 16 memory tokens beat full-context with 108x fewer tokens
omarsar0 · x · 2026-10-19
A Stanford/Toronto/MIT paper introduces Memorilla, a memory system that encodes document collections once and retrieves query-relevant info into just 16 memory tokens in a frozen LLM's input space. It beats Mem0 and MemoRAG on 5 of 8 benchmarks, hits 84.33% on FactKG vs 80.43% for full-context with 108x fewer tokens, and lifts TextWorld agent success from 18.6% to 24.2%.
More from coding & agent
- Claude uses Hedra MCP to auto-refit a music video into 9:16 vertical — AIandDesign · 2026-10-19
- Claude writes 90% of Claude Code's code; merged PRs doubled while reverts fell 40% — PrajwalTomar_ · 2026-10-19
- ESP32 + Claude Code tutorial: firmware, server setup and MCP integration in 2 minutes — alexcovo_eth · 2026-10-19
- 4 Gemini 4 Argon agents team up for 27 hours to beat Minecraft's Ender Dragon — robertarail · 2026-10-19
- Designing a non-bugged activation-read probe: numerics must not shift under the same prompt — brandon_xyzw · 2026-10-19
- OpenCode 2.0 ships: major release for the open-source terminal coding agent — RelevantEmergency707 · 2026-10-19