QUARTERMASTER

A catalog of every file on the machine, keyed by SHA-256, with every place each one has lived. On top of the catalog: duplicate detection that quarantines instead of deleting, tags, labels and ratings, local-model AI tagging, and macOS Finder-tag sync. A CLI does the work; a read-only web UI, a native Mac app and Grafana dashboards read it. Athenaeum, a Markdown-native knowledge base compiled by an LLM, lives in the same repo.

• > Content-addressed catalog: one row per SHA-256, every path it has been seen at
• > Safe dedup — protected paths are refused, and the default action is a 120-day quarantine, not a delete
• > AI tagging runs against a local Ollama model; nothing is sent off the machine
• > Internal tool — no public domain

> what it does

QUARTERMASTER catalogs every file on a machine by its SHA-256 and remembers every path each one has been seen at. The catalog lives in Postgres; a qm CLI ingests, verifies, finds duplicates, and attaches metadata.

Three ways to describe a file sit on top: nested tags, flat labels, and MPAA-style ratings with a history of who rated what. Tagging can run unattended against a local Ollama model overnight, and macOS Finder tags are imported. A read-only web UI, a native Mac app, an egui light table and a set of Grafana dashboards all read the same database.

ATHENAEUM ships from the same repo: a personal knowledge base that treats the LLM as a compiler rather than a search engine. Notes are captured as Markdown, compiled into an Obsidian-compatible vault, and queried without a retrieval index. It is still at the architecture-draft stage.

> what’s inside

• >Content-addressed catalog with location history, plus a verify pass that catches drift.
• >Duplicate handling that refuses protected paths and quarantines for 120 days instead of deleting.
• >Local AI tagging through Ollama for text and images — nothing leaves the machine.
• >Finder-tag import and a daily sweep of ~/Downloads.
• >Grafana dashboards straight over the live catalog.

> why it exists

Years of files spread across disks end up as copies of copies with no memory of where they came from. Hashing everything once turns “which of these can I delete” into a query, and quarantining instead of deleting makes the answer safe to act on.

Internal tool — no public domain.