Foundational AI Research

Building the systems that let machines remember, reason, and act.

We are a research lab working at the foundations of intelligence — hyperbolic memory, multi-agent harnesses, real-time voice, and geometric deep learning. Quietly, deliberately, and from first principles.

Research Directions

Hyperbolic Memory Systems

01

Memory embedded in curved space, where hierarchy and relation compress naturally — giving agents recall that scales with context.

Multi-Agent Harnesses

02

Control structures that let language models plan, use tools, and recover from failure reliably enough to run real work end to end.

Geometric Deep Learning

03

Models that respect the symmetry and structure of their data, learning over graphs, manifolds, and groups rather than flat vectors.

Voice

04

Low-latency, full-duplex speech systems — turning real-time conversation into a first-class interface for autonomous agents.

Thesis

Today's agents are capable but brittle. They forget, they drift, and they break the moment the world stops looking like their training data.

We think the missing pieces are structural, not just larger. Reasoning needs a harness. Memory needs the right geometry. Interaction needs to happen at the speed of speech. We build each layer with the inductive biases that make intelligence durable rather than merely impressive in a demo.

The work is early and most of it is not public yet. If these problems are the ones you think about too, we want to hear from you.