An AI lab developing the advances required for memory in models.
Our long-term work is model memory: how models retain, update, and use knowledge over time. We test that work through domain-specific models and agents in production.
What we are building toward.
We believe the next major leap in AI requires memory inside models: systems that can retain knowledge, update what they know, and reason over accumulated context without depending only on static retrieval.
Asteras exists to develop those long-term advances. Our applied work with companies gives us the data, constraints, and production feedback needed to train domain-specific models and agents that move that agenda forward.
The people behind the work.
A small interdisciplinary team across machine learning, systems engineering, product, and applied AI — Olympiad-level mathematics, a seasoned tech executive, production AI from Google, Amazon, and Nubank, and MIT-grade research — focused on the model-memory problem. The team works end-to-end: understanding domain workflows, adapting models, building agents, evaluating behavior, and deploying systems that can be operated in the real world. Based in Brazil.
How we work.
Three working principles we try to actually live up to.
First principles
We resist accepted answers. When approaching a problem, we take it apart until what remains is genuinely irreducible — and rebuild from there.
Long-term model memory
We work on the capabilities models need to remember, revise, and use knowledge over time. Applied systems are how we pressure-test that work.
Research that reaches use
We connect long-horizon technical work to production constraints, so the models we build are not only interesting, but measurable, operable, and useful.
Want to work together?
Reach out about engagements, partnerships, or just to say hello.