Mnevra

Most “second brain” tools are filing cabinets: you do the remembering, they do the storing. Mnevra inverts it. You talk — text or voice — and it extracts the people, places, and events from what you said, deduplicates them into a knowledge graph, links them with typed relationships, and then does the thing no notes app does: it notices what you haven’t told it yet, and asks.

Every morning, one question

The flagship feature is an auto-biography. The system continuously scores your accumulated memories against life chapters — early life, career, relationships, places — and identifies the gaps. Each morning, Neve (the resident interviewer persona) sends one push notification with one question. Behind that gentleness is a strict little state machine: each gap tracks engagement, attempts, and deferrals; two ignored nudges and it backs off; an engaged-but-unfinished chapter gets a fortnight’s cooldown. An interview never digs more than three turns on one topic — the bar is “could write a paragraph,” not an interrogation. When you want the book, it’s generated chapter by chapter through batch APIs and rendered to PDF in your choice of literary voice — Bryson, Didion, Hemingway — with annotations feeding corrections back into the next edition.

Here’s what that actually produces: Soundtrack to a Life — a real chapter Mnevra interviewed out of me, one morning question at a time. I picked the Creative writing style, so the flourishes are the machine’s; the memories are mine.

The log is the truth

The architectural bet: every message, tool call, and result lands in an immutable log, and the entire knowledge graph is treated as derived state. When a better model arrives or an extraction prompt improves, the log replays and the graph rebuilds. In an AI system, “what did the machine conclude?” matters as much as “what did I say?” — this design keeps both, forever, audited.

Retrieval, scarred into shape

Memory recall runs on pgvector embeddings — but pure semantic search earned a scar: during one biography run it missed a friend and a medical appointment simply because neither was semantically close to the chapter topic. Retrieval is now two-bucket — one bucket anchored to the topic by keyword and cosine similarity, a second sweeping in everything high-relevance regardless of similarity — so the biographer never writes around a fact it technically had. Voice runs on a realtime API with server-side tool execution and per-modality cost accounting, because audio tokens bill differently and a personal AI that surprises you on price stops being used.

Dogfood, dated and numbered

Mnevra’s primary test subject is my own life — hundreds of memories, a couple hundred entities, ten thousand embeddings — and the commit log reads like a diary of being your own user: feedback report #5 is literally “Neve re-asked how I met my partner.” The notification cadence got retuned from real annoyance, the voice personas got re-cast (Alfred is now a proper British valet), and the interview quality improved the way everything here improves: by being lived with, daily, until the rough edges file themselves smooth.