Inventory the memory of an agent you use

The work is classification, so the code stays small. src/locate.py finds where Claude Code keeps each kind of memory for one project, using only file stats, and src/tables.py checks a table for the three mistakes the review is about: an episode filed as a fact, an agent writing its own procedures with nobody reviewing them, and client data in a firm-wide store. The decision that matters most is the column written_by. Claude Code's own docs split memory into what you write (CLAUDE.md, rules, skills) and what Claude writes (auto memory), and that split is the procedural versus semantic split you need in the fleet.