A self-hosted memory engine for your AI assistant. Every event is stored in an append-only PostgreSQL log on your server, consolidated nightly into searchable memories with a temporal knowledge graph, and served to any MCP client. The SQL dump is the export; the AI layer is a swappable adapter.

What does it cost to run?

About zero beyond a small VPS: consolidation runs on the AI backend you pick (Claude CLI, an API key, or any open-source model via Ollama); embeddings and transcription run locally on CPU.

A Linux VPS with 2 vCPU and 4 GB RAM runs the core. Local bricks scale with the models you pick — lightweight embeddings fit in ~500 MB; large word-aligned transcription models want a few GB more. All CPU, at personal volumes.

The same MIT code, packaged for a 15-minute install: a versioned tested release, the illustrated AI-first install walkthrough, and ready-to-import iOS capture shortcuts. One-time payment, free re-downloads.

Souvenance — a living memory you own

core banks

SIX CELLS. ONE MEMORY.

CELL·01

╔══╗ ║██║ APPEND ╚══╝ ONLY

OWN

Every event lands in an append-only log in PostgreSQL. On your VPS.

No update, no delete — enforced by the database itself. A better engine in 2028 can re-read your whole life. The SQL dump is the export.

CELL·02

▄▄▄▄▄▄ z Z z ▀▀▀▀▀▀

SLEEP

Every night at 3:30, the machine consolidates — like a brain.

Extraction, importance scoring, contradiction detection, active forgetting. A weekly REM pass recombines memories across life domains.

CELL·03

1994──╮ 2019──┼──╮ NOW ──╯ ╰──

TIME

A temporal knowledge graph: what was true, and when.

"What did I believe last year?" is a query, not a reconstruction. Facts carry validity intervals; contradicted edges close but never vanish.

CELL·04

MIC ▸▸ ██ IMG ▸▸ ██ TXT ▸▸ ██

CAPTURE

Two taps on iOS: a WhatsApp voice note, a photo, a PDF — archived forever.

Word-aligned transcription overnight. Originals never modified. Memory is never a chore — it's the byproduct of tools you already use.

CELL·05

▶ ───── your own voice, 2026

REPLAY

Ask, and get the summary plus the original audio — one tap, your voice from that morning.

Doubts included. A recorded autobiography written in the present tense, before your future self rewrites the past.

CELL·06

[AI]⇄[AI]⇄[AI] swap any time

SWAP

The AI layer is an adapter: a CLI, an API key, or any local model via Ollama.

If the AI layer breaks, events keep accumulating and consolidation catches up later. Pause, never loss. No lab holds your context.

signal path

DAY CAPTURES. NIGHT REMEMBERS.

YOU

2 taps / 2 min

EVENT LOG

append-only · yours

03:30 ☾

consolidation

MEMORY

vectors + time graph

YOUR AI

any MCP client

You live your day. The machine files it at night. Next morning, ask your assistant "what did I think about X last year?" — and it answers from your archive, with dates, confidence levels, and a link to the original recording.

diff against the world

WHY THIS EXISTS

--- their_memory.txt +++ souvenance.me

- Native AI memory: written without you, stored at the lab, non-portable. Their lock-in.

+ Yours: auditable SQL on your server. Leave any provider, keep every memory.

- Legacy apps: memory for others, after you. A printed book, a posthumous profile.

+ Memory for you, alive: query it daily. It still becomes a dynamic autobiography.

- Agents that act first, remember badly, and notify you constantly.

+ The silence rule: it never calls you back. Deep when you come; silent when you leave.

acquire

TWO WAYS IN.

THE ENGINE

Complete MIT-licensed engine — nothing held back
install.sh, docs, tests, CI
For self-hosters at ease with a reverse proxy and an afternoon

OPEN SOURCEgithub.com/Pensine/souvenance.me

THE KIT

$69

Illustrated install walkthrough, written to be followed by your AI — "paste, say install it"
iOS capture shortcuts, ready to import + every pitfall documented
Versioned, pinned, tested release · free re-downloads
One-time. No subscription. You fund the project.

GET THE KITinstant download