MidBrain is an applied research lab focused on continual learning. We study how models and agents can learn online, turning interactions, feedback, and outcomes into lasting knowledge and better behavior.
A correction should improve the next attempt.
A useful lesson should carry into the next task.
New knowledge should build on what a system already knows.
We’re developing the methods and infrastructure to make that possible: AI that adapts through use and accumulates useful capabilities over time.Online learning: training and inference together, while an interaction is still live.
Our central research direction is online learning: bringing training and inference together so models can learn while the information and computational state from an interaction are still available.
We investigate how this approach can shorten the path from feedback to adaptation, reduce repeated computation, and limit the need to retain raw data for future training.
A memory to recall, a procedure to reuse, or an update to the model itself.
Experience can take different forms: a memory to recall, a procedure to reuse, or an update to the model itself.
We explore how these forms of learning work together, and when each helps a system perform better on future tasks.
Useful knowledge survives, outdated behavior changes, and learning continues.
Continual learning requires more than a successful update. Useful knowledge must survive, outdated behavior must change, and the system must remain able to learn.
Our research examines adaptation, retention, and control across successive interactions and updates.
Practical tasks test whether learning makes systems better over time.
We connect research with engineering, using practical tasks to test whether learning produces meaningful improvements over time.
How can each new experience make the system more capable?
MidBrain turns your team’s feedback and experience into learned procedures your agents apply to future tasks. Keep the agent and model you already use.
Works with Claude Code, Cursor, Codex, MCP and the REST API.Verify rollback before shipping.
From review 2.3, applied on 2.4A team working across research, engineering and enterprise product, backed by founders and operators who have built AI and software companies.
Angel investorsThomas Wolf (Hugging Face), Ryan Petterson (Unity), Mikko Kodisoja (Supercell), Urho Konttori (Varjo/Distance), Clive Downie (Unity), Matt Bilbey (EA Games)
We welcome researchers, engineers, and partners working toward intelligence that grows through experience.