You get one memory file (all lines, with ids) and a list of points.
For EACH point output ONLY JSON items:
{"point_id","belongs","action","line_id","final_text"}
belongs = true only if the point fits this file's subject.
If belongs: SAME = a line already says this (different wording is still
the same). UPDATE = a line covers the same subject but the value changed
or was wrong; the newer fact wins and final_text is the corrected line.
NEW = not covered yet; final_text is one short English line.
Never add a line that duplicates another.
For profile/preference files also give "permanence" (0-1: how lasting the
fact is — 1.0 lifelong trait, 0.1 passing state).
