Zeroset's research team is building the foundations for AI that learns continually. We're designing the models and methods behind Nebula's state layer and our Memory Models research direction. systems that accumulate understanding over time, reason over a persistent record of a user's goals and workflows, and surface the right task before it's asked for. We want to dissolve the boundary between a model and the state it operates on.
This is foundational, open-ended research. The interfaces between models and their state don't have settled answers yet, and we're betting that the people who define them will shape what the next generation of AI can do.
What you’ll work on
- Architectures for memory: graph representations, retrieval, hierarchical context, long-horizon reasoning over state.
- Training signal from traces: how a structured workflow record becomes data a model can learn from.
- Architectures for continual learning and proactive surfacing.
- Benchmarks. Help shape the evaluations the work gets measured against.
You should have
- Background in some subset of: representation learning, retrieval, graph models, sequence models, RL, LLM post-training, agent systems.
- The habit of reasoning from first principles instead of the last paper you read.
- Intellectual taste, high autonomy, a bias to disagree.
Bonus
- State-space model architecture, dynamic graphs, memory-augmented models, long-context, process mining, or graph neural networks.
- Experience designing benchmarks.
None of these are requirements. We believe someone with strong foundations can adapt to this work. If that’s you, and you want to work on the frontier of model-state interfaces, reach out.