A visual tour How it works
Your AI's long-term memory

How your AI remembers you.

Every chat leaves a trace. Here's where it goes, what happens to it, and how it comes back — in pictures.

YOU typing to your AI AI ASSISTANT Claude · Gemini · Codex TRANSCRIPT every message on disk HELPER watches the file hi 0.42… OLLAMA words → numbers QDRANT memory database 32,162 moments returned on request next chat asks Everything lives on your computer. LOCAL · PRIVATE · OFFLINE-CAPABLE
The whole system in one picture
What's in there

32 thousand moments, and counting.

Every chat leaves behind three kinds of memory.

32,162

saved moments

across 15 memory collections

Your words

what you asked

AI replies

what it answered

Session summaries

how it went

How a memory is made

Five steps, every message.

Step 01

You chat. The AI writes a log.

Every message is saved to a transcript file as it happens.

Step 02

A helper watches the log.

Quiet background worker. Wakes when the file changes.

hi 0.42 0.78
Step 03

Words become numbers.

Meaning gets turned into a fingerprint of numbers.

Step 04

Fingerprint saved to the database.

Duplicate messages are skipped — it knows if it already has one.

Step 05

Searchable by meaning.

Not keywords — topic. Ask about "payments" and you'll find chats about invoices too.

When you start a new chat

It remembers before you say hello.

The moment a new chat opens, it quietly searches for relevant past moments — and hands them to the AI before you type a thing.

NEW CHAT just opened asks DATABASE searches by topic returns top 5 "the auth bug from tuesday…" "how we structured the API…" "your preference for TS…" "the stripe integration…" "Ready. I remember."

Context, before you type.

Connected assistants

Every AI on this machine shares it.

Four assistants, one shared memory. Remember something with one, recall it with any.

Claude

Terminal · IDE

Gemini

Terminal CLI

Codex

OpenAI CLI

OpenCode

Open-source

Privacy

All local. Nothing leaves.

Your machine, your memory.

The AI runs the model locally. The database lives on your disk. No cloud uploads, no API sends, no outside eyes.

Built to keep working.

Survives crashes, reboots, force-quits and bugs. Self-healing by design.

Read the recovery story →