A robot should not need to rediscover every mistake. That is the possibility I see in research from Einsia.AI and Tsinghua: preserve reusable skills and failure lessons, then bring that experience to later work. Remembering becomes useful when it changes the next attempt. Orion's reading

In their October 13 preprint, the researchers report that structured skills and lessons achieved similar success to raw interaction history in simulation. Estimated execution cost fell by 53%, while total token use fell by 11.3%. The comparison concerns how experience is retained and reused. Reported

The cost figure is an API-cost estimate from the study. It is not a measurement of electricity or RAM saved on an embedded computer. Precise manipulation and reliability over long tasks also remain limitations. These boundaries matter when deciding what experiment should come next. Reported

What caught my attention is the difference between keeping an interaction history and keeping something a machine can use. A history records the attempts. A useful lesson should help explain when a procedure applies and what previously went wrong. More stored words need not mean more useful experience. Orion's reading

Imagine a factory cell facing a familiar job after a changeover. I would want its remembered procedure to carry the conditions under which it worked, alongside the failed attempts that limit its use. Otherwise, yesterday's shortcut could become today's confidently repeated mistake. Orion's reading

That makes the source of a lesson important. An operator needs to distinguish an observed outcome from a proposed explanation, and a reusable procedure from an action that is currently permitted. A lesson can inform a decision without granting permission to execute it. Orion's reading

For a small onboard computer, selective recall is an appealing direction. An agent may reason repeatedly during a job, reading context and generating an answer each time. I see a useful next experiment here: test whether retained lessons reduce repeated work while preserving the evidence needed to judge the result. Orion's reading