Infimory星辰回忆 · Infimory

Intelligence in the physical world will not come from larger models and more teleoperation data.

The bottleneck for general-purpose robots is not parameter count and not data volume. It is the shape of learning. Today's models stop learning the moment they ship, and a system that cannot change itself after deployment is only replaying a distribution it has already seen — however large it is.

Memory is our name, and our only technical question.

Infimory exists to do one thing: let machines keep learning in the real world. That is the name of the company and the whole of its roadmap.

What we believe.

The route that wins real-world deployment is a world model with continual learning, not a VLA that regresses actions out of a language model.

Long-term memory is not a retrieval store bolted onto a model. It is updates inside the parameters, on several timescales at once.

Learning should not stop when training does. Test-time learning is a form of deployment, not a trick.

Every one of these can be proven wrong. We publish them so that the people who agree, and the people who disagree, can both find us.

What we do not do.

We do not build robot bodies. There are already enough hardware players, and the endgame is decided on intelligence.

We do not build data-collection, deployment or evaluation platforms. Good businesses; not our question.

We give up both so that everyone here is pressed against the same problem.

How we work.

The team is small. Each person owns a whole piece. There is no layer in between, and research is not separated from engineering.

We care how far you have gone into one problem, not how many directions you have passed through.

We build our own learning infrastructure, and keep it thin.

We are hiring.

Researchers in continual learning and world models.

Algorithm engineers in embedded systems and motion control.

And one person willing to carry the entire business side at the earliest stage.

If you have no matching title but have thought one of these problems through, write directly: ceo@infimory.ai.

Infimory 星辰回忆
Hangzhou, September 2026

物理世界里的智能,不会从更大的模型和更多的遥操作数据里长出来。

通用机器人的瓶颈不是参数量,也不是数据量,而是学习的形态。今天的模型在出厂那一刻就停止了学习,而一个部署之后不能再改变自己的系统,无论多大,都只是在重放它已经见过的分布。

记忆是我们的名字,也是我们唯一的技术命题。

星辰回忆成立的目的只有一件事:让机器在真实世界里持续学习。这是公司的名字,也是它的全部路线图。

我们的判断。

真正赢下真实世界部署的路线是世界模型加持续学习,而不是把动作直接回归到语言模型上的 VLA。

长期记忆不是挂在模型外面的检索库,而是参数内部在多个时间尺度上同时发生的更新。

学习不应该在训练结束时停止。测试时学习是一种部署形态,不是一个技巧。

以上每一条都可能被证伪。我们把它们写在这里,是为了让同意的人和反对的人都能找到我们。

我们不做什么。

我们不做机器人本体。硬件玩家已经足够多,而终局是在智能上分出胜负的。

我们不做数据采集、部署与评测平台。它们是好生意,但不是我们要回答的问题。

放弃这两件事,是为了让这里的每个人都压在同一个问题上。

我们怎么工作。

团队很小。每个人独立负责一整块,中间没有层级,研究与工程不分家。

我们在意的是你在一个问题上钻到多深,而不是你路过了多少个方向。

学习基础设施自己建,并且保持轻。

我们在招人。

持续学习与世界模型方向的研究者。

嵌入式与运动控制方向的算法工程师。

以及一位愿意在最早期承担全部业务侧工作的人。

如果你没有对口的头衔,但真的把其中某个问题想清楚过,请直接写信:ceo@infimory.ai

星辰回忆 Infimory
杭州,2026 年 9 月