Toward AI with faithful memory.
Our research agenda is organized around a single conviction: that the most valuable AI systems will be the ones that remember accurately, reason transparently, and stay reliable as they grow more capable.
Memory & recall
We study how learned systems store, retrieve, and update knowledge — and how to make recall faithful to evidence rather than to plausibility.
- Grounded retrieval and citation as a first-class capability
- Persistent, editable memory that resists drift over time
- Calibrated uncertainty — models that know what they don't know
Interpretability
Understanding the internal structure of models so their behavior can be inspected and explained, not just observed from the outside.
- Mechanistic accounts of how representations form
- Tools that surface why a model produced a given answer
- Evaluations that test understanding, not pattern-matching
Alignment & safety
Designing systems that behave reliably under real-world pressure, with oversight that scales as capability grows.
- Robustness to distribution shift and adversarial input
- Scalable oversight and human-in-the-loop control
- Graceful, legible failure modes
Grounded perception
Connecting language to the world — text, images, and structured evidence — so understanding is anchored rather than imagined.
- Multimodal models that read documents and scenes accurately
- Evidence-linked generation over hallucinated detail
- Faithful summarization of long, messy source material
Environmental stewardship
Frontier AI consumes enormous power and water to cool its hardware, straining the grids and communities where data centers live. We treat that cost as a first-class design constraint.
- Efficient models and inference that do more with less compute
- On-device, local-first execution to cut data-center load
- Transparency about the energy and water footprint of our systems
Information security
Cloud models send personal data across the internet to be processed and, often, retained for training. We build so sensitive information never has to leave the device to be useful.
- Local-first processing that keeps personal data on your hardware
- Credentials and secrets held in the system keychain, never the cloud
- No training on your private data without explicit, revocable consent
We publish what advances understanding, and hold back what could cause harm.
As our work matures, papers, notes, and open tools will appear here. If you'd like to collaborate or follow along, we'd love to hear from you.