Most memoir software assumes the storyteller is the user. That sounds reasonable until you picture who the storyteller usually is.
An 84-year-old grandfather doesn't want onboarding. He doesn't want account recovery, or another app, or a workflow to learn. He wants someone to ask a good question and actually listen.
MemoirQuill, an AI memoir tool Novadiem built from that one observation, is designed around it. The subject never touches the system. They never create an account, never learn an interface, never manage a file. They sit down and tell a story. Everything else is someone else's job.
This piece walks through how a spoken conversation becomes a memoir chapter, and why almost every decision underneath points back to a single rule: protect the source.
A different architecture
Most software pulls the subject into the system. MemoirQuill does the opposite. It sends the system to where the subject already is.
A family member sets up the interview once. Background notes, interviewer instructions, and a session plan are prepared ahead of time inside an AI environment the family already trusts, like ChatGPT or a Gemini Gem. When it's time, the subject opens the project, says "begin," and talks. The assistant runs the interview. It asks the follow-up questions and chases a memory when one surfaces, working through a life one chapter at a time.
Afterward, the transcript is exported and pasted into MemoirQuill. That paste is the only bridge.
MemoirQuill never joins the live conversation. It doesn't stream audio or watch the interview happen. It receives finished text, nothing more. The family's trusted environment stays the interview room, and the thing we build on top never gets between two people talking.
The transcript is the asset
Every imported interview becomes a SessionDocument, the canonical record that every future chapter, summary, and timeline is derived from. One rule holds it together: the original transcript is immutable. It is written once and never touched again.
The raw words (messy, imperfect, human) are never edited or overwritten. Everything downstream is a view of them: the chapters, the relationship graph, each generated paragraph.
Most systems treat raw input as scaffolding: useful until the output is built, then discarded. We inverted that. Prose can be regenerated a hundred times; a memory spoken by someone who has since passed away cannot be regenerated even once. So the transcript is the asset, and everything else is a view of it.
From conversation to chapter
Once imported, the transcript moves through a deliberate pipeline:
normalize → entity_extract → chapter_map → retelling_detect →
contradiction_detect → structured_clean → prose_draft → review_flag
Each stage does one job and can be retried on its own. The whole chain is resumable, every LLM call flows through a single controlled function, and a failed model call is an operational event the system recovers from, never a data-loss event. The source is already safe upstream, so a stage can break and rerun without ever putting the original words at risk.
Two gates are wired into the chain on purpose:
- Coverage gate: a chapter with too little source material is flagged
thin_sourceinstead of padded with filler. - Safety gate: sensitive, distressing, or incoherent content is held for a human to review before anything is published.
The goal was never maximum text. It was maximum fidelity to a person's actual life, and sometimes fidelity means writing less.
Provenance over polish
Family stories are rarely tidy. Memories contradict each other, dates drift, the same event comes out three different ways across three sessions.
MemoirQuill doesn't quietly pick a winner. If one interview says Grandpa left Montreal in 1958 and another says 1960, both dates stay on the record and the system flags the conflict for the family to see.
Provenance runs all the way down. The prose is generated as an ordered set of passages, and each one carries the IDs of the exact transcript fragments it drew from, so any line in the finished memoir traces back to specific words spoken in a specific session. The family can always return to the source, and the source never moved.
Consent is a chokepoint, not a checkbox
Before any transcript is processed, two conditions have to be true at a single point in the code:
- The administrator has acknowledged the data-flow and AI-processing disclosure.
- The account has an active subscription.
Both checks live in one place, so there is no second path that quietly skips them. Reading existing content is never gated, and cancelling a subscription never deletes a memory. The system may stop making new work, but it never locks you out of what's already there. For a product built on other people's life stories, that is a core design decision, not fine print.
The memoir is only the first view
A finished chapter is one view into a life. Under the prose, MemoirQuill already keeps a structured record of the people, places, events, and stories the subject described, each one tied back to the words that produced it.
That structure is where this goes next. A memoir is something you read once; a grounded archive is something you can ask questions of. One day a granddaughter should be able to ask "what did Grandpa say about his first job?" or "did Grandma ever talk about moving across the country?" and get an answer drawn from the original conversations, provenance intact, instead of a generic AI summary.
The timing isn't an accident. Consumer voice assistants only recently got good enough to run a warm, structured interview from a written brief. That single capability is what makes zero-friction capture possible now, and it's why the memoir is the first interface to a deeper family archive rather than the end of the road.