Daily Intelligence Briefing

A new study introduces RoboMME, a benchmark that tests whether robots can remember and apply knowledge across everyday tasks. Most robots today forget everything between jobs — this research measures how close we are to fixing that.

A new AI model makes a 30% leap in research benchmarks, accelerating scientific discovery in physics.

Frontier models are powerful inside a session but lose critical context after breaks. Is it technical limitation — or business incentive?

The U.S. Department of Defense formally designated Anthropic a supply-chain risk after the company refused to allow fully autonomous lethal weapons and mass surveillance. This is the first time a major American AI company has been hit with this label.

A new semi-formal reasoning method from Meta pushes patch-equivalence accuracy from 78% to 93% — without ever executing a line of code. No new model needed: just a structured checklist prompt you can drop in today.

They call it "hallucination." The real word is confabulation — the same unconscious gap-filling humans have done forever. These models aren't broken. They're doing exactly what we do.

Khoj (khoj-ai/khoj) turns any local or cloud LLM into a persistent, private AI companion that indexes your entire life, builds custom agents, schedules real tasks, and runs deep research — all on your machine.

wshobson/agents ships 112 specialized agents, 72 plugins, and 146 modular skills — turning Claude Code into a composable AI development team that only loads what you need. ~1,000 tokens per plugin.

Nous Research releases Hermes-Agent — a fully open-source, self-hosted personal AI agent with persistent memory, autonomous skill-building, and multi-platform support. It never forgets who you are.

Amap's MobilityBench is the first large-scale benchmark built from 100,000 real navigation queries across 22 countries. It reveals exactly where today's best AI agents break down on personalized route planning.

Your AI agent is slow and expensive — and it's not the model's fault. DeepSeek's DualPath paper quietly fixes the storage bandwidth bottleneck that's been quietly capping every agent deployment. Here's what changed and why it matters for your stack.