Recall is not change
Memory can help an agent find an old fact. It does not decide whether one correction should become future guidance, which tasks deserve it, or whether the result improved. Meraki focuses on that missing change layer.
Prefer the smallest honest rule
A useful proposal should explain the observed miss without becoming a theory of the whole person. Narrow rules are easier to review, test, rescope, and remove.
Test the contribution
A later result is evidence only when it can be compared. Meraki uses baseline, raw-memory, scoped-pack, and targeted-removal runs to ask whether the guidance itself mattered.
Learning needs an undo button
A system that can change but cannot recover is not trustworthy. Every accepted change needs provenance, a version, and a route back.