Blog
Notes on what we are working on and what we have found. Claims carry a number or a citation.
Why we built Continual-ARC.
We ran three open agent harnesses through a stream of ARC tasks that keep coming back with fresh inputs, once with their memory on and once with it wiped before every instance. Their memory saved nothing we could measure, because left to themselves they kept almost nothing about a task. Told to keep one tested program per task, all three did far better.
Read the postAll posts
We ran three open agent harnesses through a stream of ARC tasks that keep coming back with fresh inputs, once with their memory on and once with it wiped before every instance. Their memory saved nothing we could measure, because left to themselves they kept almost nothing about a task. Told to keep one tested program per task, all three did far better.
Harnesses are how today’s agents remember anything at all. They are a point-in-time fix, and continual parametric learning will replace them.