Mnemo 🗂️

Paste anything. Find everything.

A local, AI-powered note hub. Paste any mess — text, links, code, PDFs, login blocks — and Mnemo cleans it, titles it, tags it, and files it onto a logical tree. Click or search to retrieve. Built to replace sticky notes, txt files, OneNote, and scattered bookmarks — 100% local-first, runs on consumer hardware.

Mnemo is not a hardcore password manager — use a real vault (e.g. Bitwarden) for true secrets. Mnemo gives fast, organized, lightly-protected access to notes and structured credentials, with a clean hook to add full encryption later.

Quick start

git clone https://github.com/yousaf-butt-b4y5/mnemo
cd mnemo
pip install -r requirements.txt
python -m uvicorn app:app --port 7575      # Windows: .\run.ps1

Then open http://localhost:7575. Defaults work out of the box with a local Ollama model (gemma3:4b); no cloud key required.

• Paste into the top box → Capture (or Ctrl+Enter).
• Drag-drop a file or PDF anywhere → it's read and filed.
• Search instantly (FTS5). Click a tree node to filter.
• Secrets are masked; click Reveal → eye-toggle → copy (clipboard auto-clears in 20s).

How it organizes

1. Secret detection (rule-based, local) runs first — passwords, API keys, SSH keys, tokens, login blocks. Detected secrets are stored locally, masked, and never sent to any LLM.
2. Non-secret content is organized by local Ollama (gemma3:4b, ~3s), falling back to OpenRouter (personal vault only), falling back to a deterministic offline rule-based classifier. It always works.

Vaults (personal vs work)

• personal — your life. May use cloud AI (OpenRouter) as a fallback.
• work — for company/sensitive data. Local-LLM-only: that vault's text never touches the cloud. Switch with the header toggle — a banner shows which vault is active.

Each vault is a separate SQLite file in data/. Keep a work vault only on trusted/managed storage and never sync it to personal cloud.

Talking to other tools

• GET /mnemo/search?q=... — a scoped surface another app (e.g. a personal AI workspace) can call to find a note. Returns non-secret entries only; secrets are excluded entirely and can never be pulled this way.

Convenience scripts (Windows)

Script	What it does
run.ps1	Start Mnemo (idempotent — won't double-start).
start-hidden.vbs	Start Mnemo with no console window (used by autostart).
install-autostart.ps1	Run once → Mnemo starts hidden at every logon. Undo by deleting the Startup .lnk.
backup.ps1	Copy both vault DBs to backups\<timestamp>\ (keeps last 20).
tests/smoke.py	Re-runnable health check (detection + HTTP). Exit 0 = all good.

Config

.env holds your live settings (gitignored — copy .env.example to start). Add an OpenRouter key to enable cloud fallback for the personal vault; leave it blank to stay 100% local.

Setting	Default
Port	7575
Local model	gemma3:4b (Ollama)
Cloud fallback	OpenRouter gemini-2.5-flash-lite (personal vault only)
Data dir	./data

Architecture

app.py                   FastAPI — capture/search/reveal/upload/scoped surface
mnemo/config.py          paths, vaults, models, port
mnemo/secrets_detect.py  rule-based secret detection (runs before any LLM)
mnemo/classifier.py      Ollama -> OpenRouter -> offline rule-based
mnemo/db.py              SQLite + FTS5; secret values never enter the search index
static/index.html        the UI (vanilla, dark, single page)

Roadmap

• App PIN/lock; optional SQLCipher full-DB encryption (the hook is reserved).
• Portable PyInstaller build (no install / no admin).
• Push to an external workspace (documents/memory) + global capture hotkey.
• Configurable large-store path; Postgres/cloud path for multi-device.

License

MIT — see LICENSE.