Trajectory lands $15M seed to build continual-learning AI layer

Trajectory is explicitly targeting one of the core gaps between today’s large models and systems that look more like general intelligence: continual learning. Most frontier models are trained once, deployed and then frozen; Trajectory wants to build the infrastructure that lets them learn safely from ongoing user interactions, potentially on a weekly or even daily cycle.

If they succeed, it changes the shape of the AI stack. Instead of monolithic models periodically replaced by bigger versions, you get a constant trickle of updates driven by real product telemetry. That could massively increase the pace at which models adapt to new domains, edge cases and adversarial behaviors, while also concentrating power in whoever controls the feedback pipelines and evaluation criteria. It’s not an accident that early customers are already using Trajectory for AI customer support agents and other high-touch applications.

For the race to AGI, this is a bet that scaling alone isn’t enough; we also need systems that learn from experience. That’s very much in line with Richard Sutton’s “bitter lesson” and more recent arguments that continual learning is essential for building superintelligent agents. Infrastructure that makes this practical for many companies will likely accelerate capabilities — and also raise the stakes for robust safety and oversight.