Robots can do backflips, run, and open doors — but ask one to revisit a room it was in yesterday, and it might still walk into a wall.

This is one of embodied AI's biggest pain points: robots can complete individual tasks but can't remember what they did before, where they've been, or what mistakes they made. On July 4, 36Kr reported that Yisheng Technology, founded by a University of Hong Kong professor's team, completed a nine-figure (hundreds of millions RMB) angel round specifically to solve this problem — building a “memory system” for robots.

Embodied AI's Next Battlefield: Memory

Over the past year, embodied AI competition has centered mainly on the “body” level: whose motors are more powerful, whose joints more flexible, whose motion control smoother, whose humanoid walks more steadily and runs faster. Unitree's robot dogs are sold globally; Tesla's Optimus has demonstrated increasingly complex movements; domestically, Agibot, UBTECH, and Xiaomi's humanoids are all iterating.

But a growing number of people realize that an agile body alone isn't enough — you also need a brain that can remember things. NVIDIA's open-sourced ASPIRE robot skill library last week was also built around the core idea of letting robots accumulate reusable skill experience — essentially a form of “skill memory.” Yisheng Technology's direction is more foundational: it aims to build long-term memory systems for robots, enabling them to remember experienced scenarios like humans do, distill lessons from successes and failures, and transfer previously learned knowledge to new tasks.

This is somewhat similar to LLMs' “context window,” but not entirely the same. An LLM's context is temporary — wiped when the conversation ends; robots need long-term memory — a door-opening trick learned today should still work when encountering a similar door next month; after breaking a glass in the kitchen yesterday, the robot should handle similar glassware more carefully next time. This kind of cross-temporal, cross-scenario experience accumulation is the core of how humans learn new skills quickly, and it's the capability robots currently lack most.

Capital Floods In as Embodied AI Enters “Hardware-Software Co-evolution”

Yisheng isn't alone. That same week, an embodied AI company founded by Tsinghua vehicle engineering alumni also completed a nine-figure angel round, focusing on automotive industry deployment; a Beijing humanoid robot company raised over 700 million RMB with state capital entering. WAIC 2026 (World AI Conference) is about to open in Shanghai, and embodied AI and Physical AI are among this year's core topics.

The flow of capital signals a trend: embodied AI competition is shifting from ‘hardware single-point breakthroughs’ to ‘hardware-software co-evolution’. You can't build a truly useful robot by piling on motors and sensors alone; you need a corresponding “brain” — memory systems, skill libraries, world models, reasoning capabilities. These software-layer capabilities are what ultimately determine whether a robot can actually do real work.

The Physical AI concept that Jensen Huang repeatedly emphasizes has this at its core: AI moving from the digital world to the physical world isn't simply about shoving an LLM into a robot body; it's about building a complete cognitive system for the physical world. Memory is the most foundational and critical piece of that system — without memory, there's no learning; without learning, robots will always need humans to re-teach them, forever stuck at the “impressive demo” stage.

Of course, robot memory systems still have a long way to go before true maturity. How do you make memory more than just “storing video” and actually abstract transferable experience? How do you avoid accumulation and interference from erroneous memories? How do you retain the most useful memories and forget unimportant details within finite storage? These are all hard problems to solve. But the direction is clear: the next set point in embodied AI isn't whose robot does the best backflips — it's whose robot can genuinely learn from experience. When robots have “memory,” they'll go from expensive toys to genuine labor that can help people get work done.