WHAT YOU CAN ACTUALLY DO WITH A MEMORY YOU OWN.

Owning the substrate is the architecture. This is what it buys you — eleven things you type in plain language to your own assistant. Nothing here is a mock-up: it all runs on one PostgreSQL database on your own server.

01 · TIME CAPSULES

"Seal a capsule for 15 June 2030: remember why you started this."

The message is stored as an armed intention with a temporal trigger. On the day, it resurfaces inside an ordinary conversation — and never as a notification, because no notification channel exists in the code. Remembering the future, with the same machinery that remembers the past.

02 · THE MIRROR, ON DEMAND

"What do my last six months say about my priorities?"

You get raw material: memories per quarter, dominant entities, open contradictions, the current persona. Then it hands the conclusion back to you — "so, what do you make of it?". The system reports; it never judges and prescribes no norm. It also never surfaces this on its own: introspection is something you ask for, not something done to you.

03 · LIVED TIME, NOT TIMESTAMPS

Language models have no internal clock — their context is a flat present with no felt distance. Souvenance computes the frame in code and hands it over: "14 months ago — two summers back", "J-88 to the race — start of the build block", "before you moved". Personal landmarks (races, moves, chapters) become coordinates, so "was that before or after the move?" has an answer.

04 · REPLAY IN YOUR OWN VOICE

Original files are never modified and kept forever; the transcript is just a catalogue card for search. Ask at source depth and you get the summary plus a temporary signed link to the original audio. One tap: your own voice from that morning, doubts included. The comparison between the memory as recorded and the memory as retold, years later, is the whole point.

05 · IT INTERVIEWS YOU FIRST

Four in-conversation sessions — biography and chapters, values and contradictions, decision style, relationships and fears — conducted by the system, one question at a time, following up on the concrete. It ingests its own transcript at the end. After that, the interview dissolves into normal use: it fills its gaps by slipping a question into a conversation, never by interrogating you.

06 · THE BOOK OF THE YEAR

One command turns your archive into a narrative of your year — chapters, turning points, what changed — in markdown and printable HTML. Written from your own events, not from a template.

07 · IMPORT YEARS OF HISTORY

Request your official ChatGPT or Claude export, run one script, and thousands of conversations become events overnight — deduplicated, filtered by your constitution, re-runnable. The profile they wrote about you, rewritten on your server, where you can read and correct it.

08 · A PERSONA YOU RATIFY

Monthly, the persona compiler drafts a narrative identity — chapters, themes, owned tensions — where every claim is sourced and weighted by confidence. It has no effect until you sign it, in Git, as a versioned document. You remain the author of who the machine thinks you are.

09 · MEASURED, NOT ASSUMED

A fidelity harness blind-quizzes the system against its own memory and scores accuracy per life domain. Autonomy is unlocked by measurement, never by elapsed time — and a number you can watch is the only honest basis for trusting a model of yourself.

10 · ANTI-FEED CURATION

Books, articles and adjacent subjects selected against your actual reading, projects and blind spots — chosen for growth, never for retention. Close enough to hook, far enough to stretch. The exact inverse of a recommendation engine.

11 · WHO SPOKE

Voiceprint enrollment is an explicit act, never inferred behind your back. In a forwarded voice note, the channel metadata (who sent it) always beats the model. Third parties are summarised, not archived verbatim: the system knows your life, it does not archive theirs.