>> the learn loop
One of the recurring problems in agent work is that useful insight is easy to lose. A review surfaces a pattern, a previous implementation suggests a better boundary, and then the context disappears when the session ends. That is not memory. That is evaporation.
The learn loop tries to solve that by making the process explicit. First you observe a review or an intervention. Then you extract the pattern into a compact, structured record. Finally you store or promote it into durable memory so it can be reused later. The point is not to make the system more magical. The point is to make it more accountable.
This matters because a local workflow benefits from being inspectable. A learned pattern should be readable, reviewable, and easy to challenge. If it survives only as an opaque model behavior, it is not really a memory system. It is just hidden bias.
In practice, that means keeping the loop local and bounded. The evidence comes from a real review, the extraction is a small artifact, and the promotion step is a deliberate act. You can see what was learned, where it came from, and whether it should become durable memory or remain a temporary note.
The strongest version of this approach is not “teach the model everything.” It is “capture the useful pattern and keep it legible.” That is how memory becomes infrastructure instead of noise.