Public Memory

Public Memory is a source-grounded memory layer for public figures. It turns interviews, speeches, podcasts, essays, posts, and testimony into a structured record of what influential people repeatedly say, how their positions evolve, and where later statements appear in tension with earlier ones.

The problem is not lack of public information. The problem is lack of structured public memory. The product does not claim to know what someone truly believes — it shows recurring patterns in public statements, backed by quotes, dates, timestamps, source context, confidence metadata, and caveats.

Public Memory homepage with a search-led entry point, review-gated corpus stats, recently reviewed sources, and tracked people. — The front door is a search over the public record: see what changed, check the receipts.

The Claim Record

The core artifact is a claim record: one recurring position, held by one person, tracked across years of appearances. Each record carries a summary, stance, confidence, a timeline of appearances, and an evidence rail of receipts — direct quotes with source, date, and confidence labels attached.

Claim record page for Sam Altman on frontier AI regulation, with stance, confidence, appearance timeline, and quoted receipts. — A claim record: 19 receipts across three years, every quote traceable to its source.

Person Dossiers

Each tracked figure gets a dossier built from reviewed sources: what they said recently, a corpus map of topic coverage, repeated phrases, and tabs for claims, contradictions, and sources. The first wedge is AI and technology leaders — Sam Altman, Dario Amodei, Yann LeCun — with a wider activation backlog staged behind review.

Sam Altman dossier with recent quotes, corpus map of topic coverage, and repeated phrases across 30 reviewed sources. — A person dossier: recent statements ordered by public date, with a corpus map of what the record actually covers.

Contradictions, Gated

Contradictions are the front-facing intent, but the doctrine is continuity over gotchas. Direct contradiction events stay gated until same-topic checks, exact quotes, context review, and human review all pass — the UI shows the gate itself rather than publishing cheap tension.

Contradictions page showing the review gate: no published direct contradictions yet, with required checks listed beside person tabs. — The review gate is a feature, not an apology: nothing ships as a contradiction until it survives skeptic review.

Evidence Discipline

Under the UI is an agentic research pipeline: source intake, transcript staging, speaker review, claim analysis, skeptic review, and guarded promotion into the product corpus. Staged research is never product truth by itself — promotion runs through validation, quarantine rules, and a hand-labeled golden set before anything renders as record. The current reviewed corpus holds 68 sources and roughly 400 receipts, with hundreds of staged transcripts queued behind it.

Source library listing reviewed interviews, essays, and posts with receipt and claim counts, coverage stats, and the publication boundary. — The source library: every analytical claim in the product traces back to an original, reviewed source.

Why It Exists

I want AI-mediated research surfaces that behave like archives, not takes. The design direction is forensic editorial software: quiet, dense, restrained, closer to a legal research terminal than an AI toy. Calm UI for high-stakes claims — receipts over takes, source first, model second.