Building the memory and continual learning layer for AI agents.
Today's agents operate in short loops.
They process inputs, generate outputs, and reset.
They do not accumulate experience.
They do not adapt over time.
We believe intelligence is not just inference —
it is the ability to change through interaction.
We researched this across real environments:
Different domains. Same limitation: experience is stored, but not internalized.
Memory allows systems to recall the past.
Learning requires systems to update because of it.
Today's systems retrieve information. They do not change behavior.
To learn from experience, a system must update while interacting with the world.
Without this, memory is just storage.
SmartSearch is our first step toward this vision. A structured memory retrieval system for agents operating over long horizons. It retrieves the right experience efficiently — because without correct retrieval, learning is impossible.
We tested SmartSearch on the Linux kernel (~2GB), comparing it against a standard LLM with tool use (grep, etc.). As tasks get longer, SmartSearch keeps reasoning grounded by ranking the most relevant memories instead of expanding context. Using our index-free approach, we avoid the massive storage overhead of traditional semantic indices while maintaining stable performance across long execution chains.
SmartSearch running on Linux kernel codebase (~2GB)
We validate MidBrain in environments where learning from experience matters most.
Agents that learn and adapt over time.
AI agents that behave like real humans in simulation.
We are not building better retrieval.
We are building systems that learn from experience at runtime.
From static memory and retraining cycles
To continuous adaptation and behavior that improves through use.
We are working with a small number of design partners building long-running AI agents.