continual learning AI startups enterprise AI model training post-training

Trajectory Wants to Give AI Products a Memory. Can It Deliver?

27 May 2026

A group of former researchers from Google DeepMind, Apple, OpenAI, and Meta have launched a startup called Trajectory, which aims to build a platform enabling AI products to continuously learn and improve from real-world user interactions. The company has raised $15 million in seed funding and already works with AI-native clients like Clay and Harvey, using post-training techniques to update models as frequently as weekly based on instances where the AI falls short. While critics note this falls short of true continual learning in the traditional sense, Trajectory's founders argue it represents an early step toward AI that could eventually update in real time — every hour or even every interaction.

A new startup called Trajectory emerged this week with a founding team drawn from Google DeepMind, Apple, OpenAI, and Meta Superintelligence Labs. The pitch: a platform that lets AI products actually learn from how people use them, rather than staying frozen at whatever capability level they shipped with.

The problem they're targeting is real. Today's AI systems, however capable, stop improving the moment training ends. The model you used last Tuesday will make exactly the same mistakes next Tuesday. OpenAI, Google, and Anthropic have built increasingly powerful systems for structured domains like coding and maths, but generalising continual learning to messier real-world applications has proved stubborn. At NeurIPS in December 2025, Turing Award winner Richard Sutton made the case that continual learning is a hard prerequisite for anything resembling superintelligent AI. Not a nice-to-have. A prerequisite.

Trajectory has raised a $15 million seed round at a $115 million post-money valuation. Conviction led the round, with Bessemer Venture Partners, Radical VC, and BoxGroup alongside. Jeff Dean, chief scientist at Google DeepMind, and Stanford professor Fei-Fei Li both put money in personally. That is a credible enough set of backers to take seriously.

CEO Ronak Malde was an AI researcher at coding startup Windsurf before Google DeepMind acquired the company's top talent in a $2.4 billion deal last year. His cofounders are Arjun Karanam, formerly of Apple where he worked on Vision Pro, and Michael Elabd, who came from Google DeepMind's robotics division. Eleven people total so far.

Malde's central argument is that AI coding tools like Cursor have quietly been doing a version of this already, feeding real interaction data back into post-training cycles and shipping regular model updates. He thinks that feedback loop is a significant reason coding AI has pulled ahead of other categories, and why the big labs have scrambled to build their own coding products. Trajectory wants to export that approach to everything else.

The mechanics are straightforward enough. Rather than dropping a customer onto a generic OpenAI or Anthropic model, Trajectory starts them with an open-source model post-trained for their specific use case. Then it watches what goes wrong. For Decagon, which builds AI customer support agents, that means logging every query that gets punted to a human instead of resolved by the AI. Those failure cases feed a weekly retraining cycle. Trajectory claims the resulting models outperform frontier models on the narrow tasks that actually matter to a given business.

There is an obvious objection here, which Trajectory at least acknowledges. Coding has a clean success condition: the code runs or it doesn't. Most other domains are murkier. Karanam says part of the platform's job is helping businesses define what success even looks like for their specific product, which is a harder problem than it sounds.

Another honest objection: weekly updates with a static model between cycles is not continual learning in any rigorous sense. It is scheduled batch retraining with better data pipelines. Useful, but let's not oversell it. Elabd accepts this, framing it as a starting point rather than the destination. The stated roadmap is daily updates, then hourly, then eventually per-interaction. Whether the infrastructure required to do that at scale for enterprise customers is buildable at a reasonable cost remains an open question.

The broader commercial case rests on enterprises currently needing teams of expensive forward-deployed engineers to keep their AI stacks functional. If Trajectory's platform can reduce that dependency, there is a real market. Current customers include Clay, an enterprise sales startup, and Harvey, which does legal AI. Both are AI-native companies. Fortune 500 is the longer-term target.

The vision is genuinely interesting. The gap between that vision and a weekly retrain loop is also genuinely large. Worth watching, but the hype should be measured against what the product can actually do today.