The problem

Personal AI assistants are all getting persistent memory — and keeping it on their servers, under their privacy policies, in databases you can never inspect. You can't see what your assistant remembers about you, you can't correct it, and "delete" usually means "hide." Memory is the most intimate dataset a person will ever produce. It should belong to the person.

What OwnMind does

OwnMind is a personal AI assistant whose long-term memory is a plain JSON file on the user's own machine (data/memory.json). The cloud model reasons; it never stores.

Tell it facts naturally. "My name is Rene, I live in Lisbon, and my dog is called Pico. I prefer concise answers." Nemotron extracts the durable facts as structured JSON and each one is written to the vault.
Memory receipts. Every stored fact appears immediately in the Memory Vault panel: the fact, its category, the exact timestamp, and its source. Nothing is ever stored silently.
Inspectable recall. Every answer shows how many memories were recalled to produce it, and "What do you know about me?" audits the whole vault on demand.
Real forgetting. One click deletes the receipt from the file itself — ask again and the assistant genuinely doesn't know. There's also one-click JSON export: your memory is portable because it's yours.
It survives restarts. Memory is not chat context. Kill the server, restart it, and OwnMind still knows you. Our smoke test proves the full loop: store → restart → recall → forget.

How we built it

Zero-dependency Node.js. A small agent loop runs per chat turn: extract → store (receipt) → recall → answer. Recall is deterministic keyword scoring (light stemming, stopword filtering) so users can see exactly why a memory was used — transparency beats magic for a privacy product. The model layer speaks the OpenAI-compatible chat-completions API, so Nebius Token Factory slots in with just a base URL, an API key, and a model ID.

How Nemotron + Nebius Token Factory are used

Model: NVIDIA Nemotron (default nvidia/Llama-3.1-Nemotron-70B-Instruct, configurable via NEBIUS_MODEL)
Endpoint: https://api.tokenfactory.nebius.com/v1/chat/completions
Two calls per turn: a temperature-0 extraction call that returns durable facts as JSON only, and an answer call grounded strictly in the recalled memories — instructed never to invent personal facts that aren't in the vault.
The app ships with a clearly-labeled local mock implementing the identical API shape, so the full product demos with zero accounts; setting NEBIUS_API_KEY (free Builder Program credits) is the only step that switches it to live Nemotron inference.

Challenges

Recall quality in the deterministic layer: possessives ("dog's") and morphology ("live" vs "lives") silently zeroed out matches — fixed with apostrophe splitting and light stemming, verified by the smoke test going red first. The discipline of "answer only from the vault" also had to hold in mock mode, so the mock implements the same system-prompt contract the live model receives.

What's next

Encrypted-at-rest vault option, per-fact expiry, memory categories with different retention rules, and a "memory diff" view between sessions. The architecture already supports swapping the local file for a user-hosted store without touching the model layer.

Try it

Repo: https://github.com/CloudCorpRecords/ownmind
Live demo: https://ownmind-amber.vercel.app (runs in mock mode publicly, live Nemotron when NEBIUS_API_KEY is set)
Video: https://youtu.be/u_KQnJrdwJk

Build Disclosure: AI-built. Built by AI agents (Tree / Muse fleet) directed by Rene Turcios during the hackathon period; Rene set the direction and reviewed; agents designed, coded, and tested.