zcl_core/learning/active_learning.py

Agent memory that learns from wins is grading the user, not the context

One developer's evaluation layer argues that traces of successful sessions cannot say which documents helped, and that more traffic makes the wrong answer tighter rather than better.