Most AI agents have the memory of a goldfish. They execute a task, reset, and start the next one with a blank slate. For a robot operating a haul truck in a mine or managing a section of the electrical grid, that amnesia is a non-starter. The real world is a continuous stream of events, and a system that can't remember what happened yesterday is useless for the job tomorrow.
Metacognition, an Adelaide-based startup, is building the software to fix that. Its product, Zeos, is pitched as an operating system for intelligent agents and robots, with a core focus on persistent memory and safe action control. The company's bet is that industrial operators in sectors like mining and energy will pay for a layer that lets machines learn, reason, and remember over years, not just seconds [Startup Daily, September 2026].
The technical wedge: memory and safety
Zeos is designed to sit beneath large language models, providing the runtime where reasoning, acting, and monitoring operate as separate threads with their own context and permissions [GitHub - metacognitionai/NavGPT-3]. The company describes its systems as needing to update memory over "years or decades" while preventing unsafe or counterproductive actions [Metacognition AI]. This is the foundational technical problem Metacognition is tackling: moving from single-episode AI to lifelong, multi-agent systems that can be directed with natural language.
In October 2026, the company provided a public look at part of its stack by open-sourcing NavGPT-3, a vision-language model for robotic navigation. The model, released under the GNU Affero General Public License, is a 4B or 8B parameter model fine-tuned from Qwen3-VL and trained on a mix of navigation tasks, from point-goal navigation to outdoor driving [Forbes Australia, October 2026]. The release acts as both a research contribution and a demonstration of the kind of foundational model that could be managed by the Zeos layer.
A founding team built on research credibility
The company's most tangible asset, ahead of commercial traction, is its founding team. The trio brings deep academic and research credentials that are rare at the pre-seed stage.
| Founder | Role at Metacognition | Key Background |
|---|---|---|
| Anton van den Hengel | Co-founder, Chief Scientist | Director of Applied Science at Amazon; Director of the Centre for Augmented Reasoning at AIML; world's most cited researcher in Visual Question Answering [Anton van den Hengel’s journey from intellectual property law to computer vision pioneer - Amazon Science]. |
| Stephen Gould | Co-founder, Professor/Researcher | Professor at Australian National University; ARC Future Fellow; former researcher at Stanford and Amazon [Stephen Gould - LinkedIn, Twitter]. |
| Paul Dalby | Co-founder, Research Strategist | Former Business Development Manager at the Australian Institute for Machine Learning (AIML) [Three years at AIML - what did I learn?]. |
Van den Hengel and Gould are among Australia's most recognized AI researchers, with the former having been a chief investigator on over $80 million in research funding [Prof Anton Van Den Hengel | Researcher Profiles | Adelaide University]. This background gave them immediate credibility with deep-tech investors.
The $15 million pre-seed vote of confidence
That credibility translated into significant early capital. Metacognition raised a $10 million pre-seed round led by Main Sequence, the CSIRO-backed venture firm, in September 2026 [Startup Daily, September 2026]. Just weeks later, in October, Australian VC Airtree topped up the round with an additional $5 million, bringing the total disclosed pre-seed to $15 million [Forbes Australia, October 2026]. The speed and size of the rounds underscore investor belief in both the team and the ambitious scope of the problem.
The capital is earmarked for developing specialized AI software layers for target industrial sectors [The SaaS News]. The company is actively hiring for roles like Machine Learning Engineer, Robotics Engineer, and Research Scientist, indicating a build-out of its technical team [Ashbyhq, retrieved 2026].
Targeting the industrial floor
Metacognition's initial focus is squarely on heavy industry. The company is targeting applications in mining, energy, heavy machinery, and electricity-grid management [Startup Daily]. The use case is operators using voice or natural language to direct complex systems, such as managing an open-cut mine or grid segment through spoken commands. These are environments where the cost of failure is high, the operational timelines are long, and the value of a system that learns continuously is theoretically immense.
However, the path from simulation to deployment is the company's central challenge. One report notes that while the team has built simulations for these industrial use cases, it has not yet disclosed live deployments or customer names [ontargetish.com]. For all the technical promise, the product must first prove it can handle the messiness of real-world physical operations.
The scale test: from simulation to site
The technical breakdown of Metacognition's approach reveals both its sophistication and the hurdles ahead. The Zeos architecture, with its threaded runtime for reasoning and acting, is a logical design for managing complex, long-running tasks. Open-sourcing NavGPT-3 is a smart move to engage the research community and bootstrap ecosystem development around navigation.
The sober assessment lies in what could go wrong at scale. Industrial environments generate torrents of noisy, unstructured data. A memory system designed for "years or decades" must have near-perfect data curation and retrieval, or it risks becoming a warehouse of corrupted or irrelevant context. The safety mechanisms, which are meant to prevent unsafe actions, will face an infinite set of edge cases once outside a controlled simulation. A single critical failure in a mining or energy context could set adoption back years, regardless of the underlying research pedigree.
Metacognition's next twelve months will be defined by its ability to transition from a promising research-driven project to a deployed industrial product. The key milestone to watch is the announcement of a first commercial pilot with a named customer in its target sectors. With $15 million, top-tier founders, and a clearly defined hard problem, the company has all the ingredients for a breakthrough. The question is whether its software can remember its way through the inevitable chaos of the real world.
Sources
- [Startup Daily, September 2026] AI startup raises $10 million to build robotics operating system | https://www.startupdaily.net/topic/funding/ai-startup-raises-10-million-to-build-robotics-operating-system/
- [Forbes Australia, October 2026] Metacognition AI caps huge pre-seed raise to run in global … | https://www.forbes.com.au/news/entrepreneurs/australian-ai-lab-metacognition-caps-record-pre-seed-raise/
- [GitHub - metacognitionai/NavGPT-3] NavGPT-3: Harnessing Context in a Hierarchical Navigation Runtime | https://github.com/metacognitionai/NavGPT-3
- [Metacognition AI] Metacognition AI company site | https://metacognitionai.com/
- [Anton van den Hengel’s journey from intellectual property law to computer vision pioneer - Amazon Science] Anton van den Hengel biography | https://www.amazon.science/blog/anton-van-den-hengels-journey-from-intellectual-property-law-to-computer-vision-pioneer
- [Stephen Gould - LinkedIn, Twitter] Stephen Gould professional background | https://www.linkedin.com/in/stephen-gould-2802523/
- [Three years at AIML - what did I learn?] Paul Dalby background | https://www.linkedin.com/pulse/three-years-aiml-what-did-i-learn-paul-dalby
- [Prof Anton Van Den Hengel | Researcher Profiles | Adelaide University] Anton van den Hengel research profile | https://researchers.adelaide.edu.au/profile/anton.vandenhengel
- [The SaaS News] Metacognition plans to use capital for sector-specific AI | https://the-saas.news/
- [Ashbyhq, retrieved 2026] Metacognition AI job listings | https://jobs.ashbyhq.com/metacognitionai
- [ontargetish.com] Report on Metacognition's simulations and deployments | https://ontargetish.com/
- [Forbes Australia, October 2026] Airtree backs Aussie AI startup after its robotics model beat Alibaba, Mistral | https://www.forbes.com.au/news/innovation/airtree-backs-aussie-ai-startup-after-its-robotics-model-beat-alibaba-mistral/