Memory Curator is an AI skill for UGOS that turns a cluttered NAS photo and document library into a deduplicated, person/time/location-organized, searchable archive — asking for explicit confirmation before deleting, moving, or overwriting anything.

Inspiration

UGREEN AI NAS owners — tech-forward households and small offices — accumulate tens of thousands of photos, videos, and documents with no organization. Duplicates silently consume expensive storage, and finding anything means scrolling folders by date. Existing NAS tools sort by date only; bulk cleanup tools delete without asking.

What it does

A conversational skill that inventories a chosen folder, finds duplicates by file content, proposes album plans by person/location/time, and archives documents with extracted key facts (type, counterparty, date, amount) into a yearly index — every destructive step confirmation-gated per the Uliya standard.

How it works

The user says "organize last year's travel photos." The skill scopes the job (folder, job type, filters) and confirms it, inventories files with a one-line count, reports content duplicates for confirmation, proposes an album plan (name, photos, dimension) for confirmation, then creates the albums and summarizes. Documents follow the same scope, extract, layout-confirm, archive, index pattern.

What makes it different

Three things: (1) confirmation-first as a workflow, not a setting — the skill can never silently delete; (2) content-hash duplicate detection instead of filename matching; (3) photos and documents in one skill with a searchable document index. Pure Markdown orchestration keeps every decision auditable.

NAS capabilities used

Files: inventory listing plus content-hash duplicate detection (Steps 2-3). Albums: find by person/location/time, create and tag (Step 4). Documents: read and extract type, counterparty, date, amount (Step 5). Every step maps to a listed UGOS capability; confirmation precedes any delete, move, overwrite, or external send.

How I built it

SKILL.md orchestration workflow in the expert-agent pattern: deterministic step chaining where each step consumes the previous step's recorded output, confirmation records as workflow state, and a strict rules block (capability scope, no scripts, no executables, no sensitive data). Built with the official Uliya Skill Creation Tool.

Current status

SKILL.md validated with the official Skill Creation Tool (static-pass: 0 errors, 0 warnings) per the Uliya Skill Development Standard, covering all three job types (photos-organize, duplicates, documents-archive); references finalized.

Validation plan

Core test 1 (duplicates): seed a folder with known content-identical duplicates under different names; verify the skill reports every group with paths and sizes, and deletes only confirmed files. Core test 2 (albums): 200 mixed photos with known people and dates; verify the album plan groups correctly and albums are created only after confirmation. Edge test: an empty and an unreadable folder; verify the skill stops with a clear message and never substitutes other content.

Limits and risks

Near-duplicates (resized or recompressed photos) are reported as similar, never auto-deleted. Person identification depends on the NAS's existing face/person index. Document extraction accuracy varies with scan quality and should be spot-checked. The Round 2 demo depends on the assigned remote NAS test day.

Built With

Share this project:

Updates

Submission history