Private Memory Investigator

Inspiration

Our photographs capture moments, while invoices, receipts and documents explain what happened around them. Yet these records often sit in separate folders, making it difficult to reconstruct the complete story of an important personal event.

Private Memory Investigator was inspired by a simple question:

What if a NAS could reconstruct the story of your home from the evidence you already own, without inventing missing details?

Our goal is to transform scattered personal files into understandable, evidence-backed memories.

What it does

Private Memory Investigator is a read-only AI Skill designed for UGREEN AI NAS and its Uliya assistant.

A user can ask:

"How did our kitchen change in 2024?"

After the user authorises the relevant folders or albums, the Skill investigates photographs, invoices, receipts and other documents.

It is designed to produce a chronological report with four evidence classifications:

• DOCUMENTED: Directly supported by available evidence.
• INFERRED: Reasonably derived from evidence, with the reasoning explained.
• UNCONFIRMED: Not established by the available sources.
• CONFLICTING: Sources disagree about an event, date or amount.

Supported findings reference identifiable source files. Gaps, contradictions and incomplete searches remain visible.

The key difference: Instead of creating another photo timeline, the Skill connects visual memories with documentary evidence.

How we built it

We developed the Skill using the official UGREEN Uliya Skill Developer Kit and its documented capability catalogue.

The parent SKILL.md orchestrates specialised agents through the supported delegate --target mechanism:

• knowledge-agent: Evidence discovery and search coverage.
• photo-agent: Photograph search and metadata analysis.
• document-agent: Information extraction and document reconciliation.
• file-agent: Source-path verification when required.

The parent Skill correlates the returned results, preserves source references and assigns evidence classifications.

It does not invoke NAS CLI actions directly.

The package also contains references/report-format.md, defining the structure of the investigation report.

The workflow is designed for read-only operations and handles incomplete results, failures, cancellations and approval requests.

Challenges we ran into

Platform constraints: Uliya Skills cannot directly execute arbitrary commands or invoke CLI actions. They must delegate work to registered expert agents.

Evidence reliability: A photograph's timestamp does not necessarily prove when an event occurred. Similarly, an invoice does not prove payment.

Incomplete information: Missing metadata, conflicting documents and unavailable source references must not lead to invented conclusions.

Hardware availability: The Skill has not yet been installed or executed on a physical UGREEN NAS. Live indexing, agent responses and model routing remain to be validated.

Accomplishments that we're proud of

We completed the Round 1 Skill package and validated it using UGREEN's official local validator.

Official validator result: static-pass

• 0 errors
• 0 warnings
• 0 informational diagnostics

We also performed negative validation checks, confirming that the validator detects unknown delegation targets and missing delegation-state handling.

The Skill includes:

• Four explicit evidence classifications.
• Source tracking and uncertainty handling.
• Read-only workflow restrictions.
• Explicit delegation failure handling.
• A structured, evidence-backed report format.

Validation scope: These are static-development results. Installation, compatibility and execution on the actual NAS remain untested.

What we learned

Reliable AI must do more than produce convincing answers.

It must distinguish evidence from inference, disclose uncertainty and preserve contradictions rather than forcing every source into a coherent story.

We learned the importance of platform-compliant agent orchestration, source provenance and transparent failure handling.

We also reinforced a fundamental user experience principle:

Ask a question → investigate authorised files → receive an explainable result.

What's next for Private Memory Investigator

If selected for Round 2, we plan to validate the Skill on the assigned UGREEN AI NAS.

Our planned tests include:

1. Evidence reconstruction: Upload synthetic renovation photographs, invoices and receipts, then test whether the Skill produces a sourced timeline.
2. Contradiction detection: Introduce documents with conflicting completion dates and verify that both claims are preserved.
3. Failure handling: Test missing EXIF metadata, incomplete results and agent failures without fabricating evidence.

We will also verify file uploading, indexing, access through Uliya expert agents and actual model behaviour.

Test results will be compared against a separate ground-truth reference.

Our long-term vision: Help families rediscover the stories hidden across their personal digital archives, with traceable evidence, transparent uncertainty and user-controlled access.

Built With

• ai-agents
• markdown
• ug-document
• ug-knowledge
• ug-photo
• ugos
• ugreen-ai-nas
• ugreen-skill-developer-kit
• uliya