Australian AI Lab Metacognition Raises $10 Million in Record Pre-Seed Round

Australian AI startup Metacognition AI has raised $10 million in pre-seed funding from deep-tech investor Main Sequence, a round that is roughly 10 times the median Australian pre-seed raise recorded in 2025. The company is developing an AI “operating system” designed to give AI agents persistent memory, stronger reasoning and safeguards that can prevent models from taking unsafe or unintended actions.

The startup was founded by AI researchers Anton van den Hengel, Stephen Gould and Paul Dalby and is taking a different approach from companies building increasingly large frontier models. Instead of competing directly with OpenAI or Anthropic on foundation models, Metacognition is developing software that sits above or around existing models and controls how they interact with people, systems and physical machines. The company and Main Sequence have not disclosed its valuation.

Building an AI Layer for Safer Decisions

Metacognition’s technology is built around persistent memory and reasoning capabilities that allow AI systems to retain and use information over longer periods. The company argues that current AI agents can lose important context as conversations and tasks become more complex, creating risks when those systems are given access to sensitive accounts, enterprise software or physical machinery.

Its longer-term goal is to create a software layer that can translate high-level human instructions into actions while restricting what an AI system is actually allowed to do. That could become particularly important for robots, where a misunderstanding or malicious instruction could have physical consequences. The company wants the same basic architecture to work across enterprise systems and machines rather than tying it to a single AI model.

From Enterprise AI to Robots

Metacognition is not limiting the technology to humanoid robots. The company is initially looking at enterprise and industrial applications, including mining and electricity-grid management, where operators could give high-level instructions in natural language while the underlying system handles complex decisions and actions. The company has also built simulations involving an open-cut mine and electricity grid to explore these applications.

Robotics remains an important part of the company's longer-term strategy. Metacognition wants its software to eventually act as an operating system for robots, allowing workers to train or direct machines through ordinary language while maintaining constraints around what those machines can do. That approach could allow one AI control layer to work across different types of robotic hardware.

A Large Bet on Australian AI Research

The size of the raise stands out in Australia's early-stage technology market. Forbes Australia reported that the $10 million round is four times the country's 2025 median seed round of $2.5 million and 10 times the median pre-seed round of $1 million, according to Cut Through Venture data. While the amount remains modest compared with the multibillion-dollar rounds raised by US AI companies, it represents a substantial early bet on an Australian AI research team.

Metacognition currently has around 10 employees and plans to expand its workforce substantially over the next year. Van den Hengel said the company expects to keep raising capital because building advanced AI infrastructure remains expensive, even when a startup is not training a frontier model from scratch.

The Challenge Ahead

The company is entering a market where AI agents are increasingly being given access to business systems and tools, making reliability and control as important as raw model performance. Main Sequence partner Bill Bartee said Metacognition is targeting the limitations created by the context windows of current AI models by developing a memory and reasoning layer for longer-running tasks.

The technology remains at an early stage, and its claims about safer and more capable AI will need to be demonstrated through real-world deployments. But the funding gives Metacognition room to build beyond research prototypes and test whether an additional reasoning and control layer can make AI agents more dependable across enterprises, critical infrastructure and eventually robotics.