126-page survey from 9 universities defines Always-On personal intelligence, grades 119 systems L0-L5
新智元 · wechat · 2026-11-16
23 researchers from 9 universities (NTU, Berkeley, Stanford, Princeton, etc.) released a 126-page survey analyzing 632 papers and 119 systems to define what makes an assistant truly "always-on."
- Core definition: operational continuity (deciding autonomously when to act without requests) + state durability (carrying preferences and unfulfilled commitments across sessions) — a "continuing loop"; running in the background ≠ always-on
- Framework: sensing, memory, and action modules coupled by six directed information channels, constrained by shared conditions and implemented via model + scaffolding
- Grading: 52 systems fall short, 67 reach L3+; Muse and Claude Duet rated L3, OpenAI dots and GrokBot L4; only 5 research works hit L5 (self-evolution within user-set bounds), no products yet
- Key pain points: missed perception goes undetected; erroneous memories form self-reinforcing "memory traps"; delayed action outcomes can turn help into burden
- Open-source diagnostic tool HALO audits loop channels and self-evolution evidence via LLM review panels and A/B ablations
More from coding & agent
- 10 open-source repos to build a one-person AI content factory, from OpenCreator to Kokoro — nikola_mr64990 · 2026-11-16
- An AI install nets a client $41,500 in a month — referrals do all the marketing — aryanXmahajan · 2026-11-16
- H3 Continuity Studio 1.0: open-source ComfyUI tool for continuous video stories — rambrogi · 2026-11-16
- Google's Jaana Dogan Says Burnout Month Project Became Agent Substrate — rakyll · 2026-11-16
- Dev Builds AutoRAG Plugin for Hermes Agent to Search Chats, Drives and Browser History — Teknium · 2026-11-16
- Developers voice despair as AI code outpaces them: 'decades of work gone' — ghumare64 · 2026-11-16