AgiBot's LIFE gives a robot a day of memory, and answers from it in 80 seconds instead of nine minutes

AgiBot has released LIFE, a framework meant to let a robot keep watching and remembering between jobs, and it has open-sourced the piece that decides what is worth keeping. That piece, RoboBeats, comes in 0.8B, 2B and 4B sizes as LoRA adapters on Qwen3.5. It reads camera frames once a second and flags when a condition holds ("is the elevator door open?") or when something new happens. Only the flagged moments get written to memory. Chinese outlets reported the release on Nov. 11 Beijing time; the technical report has been on arXiv since Oct. 30.

The most telling result is in the report's real-robot test, and it concerns cost more than intelligence. With the same answering model, AgiBot's agent scored 29.5 of 30 points on questions about a day in an office. A general coding agent handed the same memory scored 28.5. It needed 554.7 seconds per answer against LIFE's 80.6, and 3.06 million input tokens against 172.8 thousand. A robot that remembers everything is affordable only if something cheap decides what to remember, and that cheap part is RoboBeats.

A day in an office, rebuilt from video

The test used footage from an A3 Ultra robot's front camera, recorded from 21:00 on Oct. 21 to 21:00 on Oct. 22 across work areas, a meeting room and a pantry. The gate and a prefilter picked 256 windows for a larger model to describe, and the stored memory kept 1,296 keyframes. The six questions included who owned a cup left after a meeting, where a jacket was last seen and when the lunch break ended. Each was asked three times. A coding agent given only the raw video, without the memory, scored 24 and took 961.5 seconds per answer.

The authors call it "a case study of limited scale," and the memory was built offline from the recording, not live on the robot.

The gate beats sampling at the same budget

On benchmarks the gate does better than the models it is built on. On StreamingBench's proactive-output task, RoboBeats-4B fired within two seconds of the right moment 64.0% of the time, against 21.2% for its Qwen3.5-4B base. The report lists Claude Fable 5 at 71.6% and Gemini 3.7 Flash at 80.0%, under those models' own published protocols. On SSS-Bench, a new set of 618 questions about 171 videos that AgiBot also released, memory chosen by RoboBeats-2B let a reader answer 60.5% correctly, against 46.5% when the same number of frames was sampled at even intervals.

RoboBeats was trained partly on Charades-STA, one of the benchmarks it is scored on. The open question is whether a gate tuned on short clips picks the right moments in a factory or a home it has never seen. With the weights public and the code under Apache 2.0, anyone with a robot can now find out.