Memory that earns its shelf space
Episodic recall, consolidation, associative memory, user context, and world models—so agents can carry useful experience forward without turning every past detail into permanent truth.
↗Frontier AI research · Thiruvananthapuram, India
We’re teaching machines to remember, notice, and behave.
XAGI Labs works on the layers that turn a clever model into a capable agent: memory, perception, computer and browser use, harnesses, safety—and, over time, models of our own.
A note on language: “shipped,” “research,” and “future” mean different things here. We label them that way on purpose.
◉
SAFETY
AT THE CORE
SAFETY
AT THE CORE
On the benches now
Intelligence is a stack.
A model can write a lovely sentence and still forget why it opened the browser. We work on the awkward, important layers in between.
Computer use that can see twice
Accessibility trees when structure is available; OCR and visual grounding when it is not. Then outcome checks to see whether the click actually worked.
↗Browser use in the real browser
Playwright and a real-Chrome bridge, accessibility scanning, cached action paths, and recovery when pages drift.
↗Harnesses that improve the whole agent
Evaluate memory, routing, task completion, and token efficiency. Explore candidates. Reject unsafe patches. Roll back when improvement is only impressive on paper.
↗Models for products—and people
Product-specific foundation models, learned multi-model orchestration, and a general-purpose model intended for public use. No invented launch date. No mystery benchmark victory.
↗Safety is not the brake pedal
It is the steering.
Capability without human authority is just a faster way to be wrong. We design for scope, restraint, inspection, reversibility, and honest limits.
Know whose world this is.
Identity, tenancy, permissions, and context should come from trusted boundaries.
Keep humans in command.
People should be able to inspect, approve, deny, mute, disarm, and reverse.
Check the outcome.
An action is not complete because an API returned 200. The real state has to change as intended.
Say what is still hard.
Research is not a product promise. A roadmap is not a benchmark. Limitations belong above the fold.
We are working toward machines that can do more—while making sure people can still say no, stop, show me, and undo that.
We do not believe AGI becomes useful by becoming mysterious. It becomes useful when it can understand context, act carefully, learn without quietly rewriting the rules, and stay accountable to the people living with it.
Ambition gets the big whiteboard. Safety gets the permanent marker.
— XAGI Labs
Read the full manifesto →Research, under real pressure
Meet ATLAS OS.
Our company operating system turns scattered signals into shared understanding—and governed action. It is where memory, agents, harnesses, and safety have to survive contact with Tuesday morning.
See what ATLAS is now →
Come make the hard parts real
We like people who ask one more question.
Research collaboration, safety work, design partnerships, press, or a thoughtful hello—we read the messages ourselves.
Write to the lab →