Inspiration
Households keep years of photos and documents on a NAS, but the assistant starts every conversation knowing nothing about them. It cannot tell who "Grandma" is or which trip "Kyoto" was, so people repeat the same context and their richest data stays unused.
What it does
NASAgent gives the NAS assistant a lasting memory of its user and uses it to answer questions about their photos and documents.
- Memory: remember, recall, correct and forget facts about the user. The memory is a set of plain Markdown notes in a folder inside the user's own personal folder, so they can read, edit and delete it.
- Photos: answer when, where and with whom questions, add keyword notes to photos, and keep a journal of places and activities.
- Documents: answer questions about a document with page citations.
Example: after "Organize my Kyoto photos", the Skill proposes journal lines, the user edits and approves them, and they are saved. Later, "When was I last in Kyoto, and with whom?" is answered from memory plus photo search, citing dates and photos.
How I built it
It is one hub Skill. Every request first runs a read-only memory preflight, then a routing table hands the request to a module. SKILL.md holds the preflight, the routing table and the core delegate commands; references/ holds one self-contained file per task, loaded only when needed.
All work goes through the registered NAS experts: file-agent (ug-filesystem, file-workflow), photo-agent (ug-photo) and document-agent (ug-document). The Skill calls no external service and names no model. Every write is shown to the user first and goes through the platform approval flow.
Challenges I ran into
- Making memory the foundation instead of one more feature: every module starts with memory loaded and hands lasting facts back at fixed points.
- Keeping a hub Skill small enough to load, which led to the layered structure above.
- Deciding what must never be stored: no secrets, no face recognition, places kept at city level.
Accomplishments that I'm proud of
The full package is 14 Markdown files and passes the official static check with no diagnostics. The memory stays the user's own readable files, not a hidden database.
What I learned
How to design within the Uliya Skill contract: orchestrating experts through delegation, handling every delegation state, and treating each step's real output as the only input to the next.
What's next for NASAgent
It has not run on a NAS yet. Next is validating every workflow on a device, including whether photo-agent can describe several photos in one delegation, and then adding further modules on top of the same memory.
Built With
- claude-code
- markdown
- ugos
- uliya-skill-kit