Memoralink Memory Studio

Memoralink Memory Studio turns family memory fragments into a dignified, source-grounded memorial draft—without inventing facts and with mandatory human review.

Comment

Inspiration

I built Memory Studio because memories should never be invented. Families should remain in control, the dignity of the deceased should be protected, and remembrance should evolve into a thoughtful digital 2.0 experience.

What it does

Memory Studio guides a contributor through four steps:

  1. Collect: Enter fictional subject details and individual memory fragments. Every fragment receives a stable source ID.
  2. Compose: Turn those fragments into a structured draft. Every substantive statement must cite at least one known source ID.
  3. Review: Compare sources and draft statements side by side. A person can edit, accept, or reject every statement.
  4. Preview: Inspect a dignified private memorial-page preview. Nothing can be published from the prototype.

Missing information is shown as an explicit open question instead of being invented. Contradictions remain visible, and confirmed direct quotations must remain verbatim.

How we built it

The prototype is a full-stack Next.js 16 application using React 19 and TypeScript. Zod schemas validate requests and structured draft responses. A server-only /api/compose boundary separates the interface from composition logic.

The public demonstration uses a deterministic local composer, allowing judges to test the complete experience without an API key, cost, or external data transfer. A guarded OpenAI Responses API adapter was also implemented behind independent server-side activation gates, a positive call budget, a single-call lock, and a cooldown.

The grounding audit rejects statements with missing or unknown source references and detects changed direct quotations. Source fragments are treated as untrusted data, so embedded prompts, commands, or role changes cannot control composition.

The project includes responsive desktop and mobile interfaces, automated workflow and failure-mode tests, and a live Vercel deployment using fictional data only.

How Codex and GPT-5.6 were used

Codex supported the project throughout Build Week: designing the architecture, implementing the source-grounding checks, writing automated tests, debugging the review and reset flows, and preparing the technical documentation.

One explicitly approved private GPT-5.6 acceptance request was run with six fictional source fragments. GPT-5.6 returned six source-linked statements, three open questions, and zero contradictions. The local schema and grounding audit passed. The provider dashboard recorded one request, 766 tokens, and USD 0.02 total cost. The adapter was immediately returned to its disabled, zero-call mock defaults after verification.

The public demo therefore remains deterministic and safe to test, while the genuine GPT-5.6 integration is implemented and verified.

Challenges

The hardest part was making the workflow simple for families while keeping every step respectful and dignified.

Other challenges included preserving exact quotations, exposing uncertainty without degrading the writing experience, treating family-provided text as untrusted input, and making mandatory human review feel like a meaningful product feature rather than an obstacle.

Accomplishments

  • Complete Collect → Compose → Review → Preview workflow
  • Stable source IDs attached to every substantive statement
  • Explicit open questions instead of invented details
  • Mandatory edit, accept, or reject controls
  • Grounding and quotation-integrity validation
  • 38 automated tests plus lint, TypeScript, and production-build verification
  • Responsive desktop and mobile experience
  • Working public demo with fictional data only
  • Verified bounded GPT-5.6 integration
  • No automatic publication or production-data access

What we learned

Trustworthy AI writing is not only about generating better prose. It is about making provenance, uncertainty, and human responsibility visible in the product itself.

Codex accelerated implementation and verification, while GPT-5.6 demonstrated that structured, source-linked composition can remain useful without silently filling gaps.

What's next

Future work includes encrypted project storage, authentication, multi-contributor review, media provenance, consent controls, and stronger adversarial testing. Any production model activation would remain server-side, bounded, auditable, and subject to explicit human approval.

Built With

Share this project:

Updates