A$10M · Metacognition + 2 more

What they're building

Metacognition is building the layer that sits *underneath* a large language model rather than on top of it. The company's argument is that today's models are text predictors with no state: they are briefed fresh every session, forget what they learned, and cannot be inspected afterwards. Metacognition's architecture separates language from knowledge representation, so the model does the talking while an explicit, persistent knowledge store does the remembering. Around that sit three capabilities the founders keep naming: memory that survives across sessions, reasoning that adapts from prior experience, and controls over what actions an agent is actually permitted to take. The first product is Zeus, which the company calls an operating system for agents. Zeus governs how models interact with machines and external systems, which is what turns a chat interface into something that can be handed a physical process. The team's own shorthand for it is 'Windows for robots': the missing software platform that lets robot hardware be trained to do useful tasks by non-experts giving simple verbal instructions and verbal feedback, rather than by an engineer writing task-specific code. The initial targets are industrial, not consumer. The company has demonstrated simulations for directing open-cut mining operations and for managing electricity grids through voice-based commands. Both are cases where the useful thing is not fluent language but an agent that accumulates context over months and whose decisions can be inspected afterwards. Safety is architectural here rather than bolted on. The company has adopted Australia's voluntary AI Safety Standard as a baseline and says it has additional internal benchmarks built into the architecture from inception, with van den Hengel stating they are 'prioritising building safety and human control into the design.'

Why this matters

The credentials behind this are unusually heavy for a pre-seed. Anton van den Hengel has an H-index of 92 across 400+ papers and 40,000+ citations, grew the Australian Institute for Machine Learning to 130 researchers, and spent four years as a Director of Machine Learning at Amazon. Stephen Gould has an H-index of 63 and 27,000+ citations, was a Principal Research Scientist at Amazon, and co-founded Sensory Networks, which Intel bought in 2013. This is not a team learning the field on the round. The thesis is also a real gap rather than a positioning exercise. Gould's framing is the sharpest version of it: 'Right now, using AI feels a bit like working with someone who's briefed fresh every meeting.' Agent memory today is mostly retrieval bolted onto a stateless model, and it degrades exactly where industrial deployments need it most — over long horizons, across sessions, under audit. Whether an explicit knowledge-representation layer beats bigger context windows is genuinely unsettled, which is the point: Main Sequence is underwriting a research bet, not a go-to-market one.