---

Hey — I'm Rohit, the one behind this.

Every assistant you've used is reactive: you ask, it answers. Vivedly is the other shape. It watches what you're actually working on, keeps a memory of how you work, and surfaces the right thing before you think to ask for it.

The hard part isn't the model. It's knowing when to speak — and, far more often, when to stay quiet. That restraint is the whole product.

The idea, concretely

You're three tabs into a bug, and the error you're staring at is one you already solved six weeks ago. A reactive assistant waits to be asked. Vivedly already knows — it saw the stack trace, it remembers the fix, and it offers it once, quietly, in the corner. If you ignore it, it learns that too.

How it's built

src/
├── proactivity/   the trigger engine — decides IF and WHEN to surface anything
├── memory/        tiered recall: RAM → SQLite → learned patterns → vector → long-term
├── learning/      lessons drawn from what you accepted and what you dismissed
├── autopilot/     multi-step actions it can carry out once you say yes
├── inference/     model routing — local first, cloud when local won't do
├── desktop/       real OS control: windows, focus, a scripting host
├── integrations/  Gmail · Slack · Notion · GitHub
├── mcp/           an MCP server, so other agents can use Vivedly's context too
├── voice/         streaming speech in, speech out
├── renderer/      the companion orb, the overlay, the suggestion widget
└── main/          Electron main process, tray, IPC surface

Memory is a hierarchy, not a vector store

Most "AI memory" is one embedding table and a similarity search. That answers what is this like — but a coworker needs what happened, when, and what did I do about it. So recall walks tiers in cost order and stops at the first one that can answer:

Tier	Holds	Reached for
1 — hot	the current session, in memory	anything about right now
2 — warm	SQLite, on disk	today, this week, this project
3 — patterns	recurring behaviour it has extracted	"you always do X after Y"
4 — vector	embedded episodes	fuzzy, semantic recall
5 — cold	long-term archive	the thing from six weeks ago

Similarity is one of five tools here, not the whole toolbox.

Inference

Local model first. Cloud only when the local one genuinely can't do the job — which means the machine stays useful on a plane, and your screen contents don't leave the device by default. Provider-agnostic: swap the backend without touching the trigger engine.

Running it

npm install
npm run dev        # launch with HMR
npm run typecheck  # TS across main, preload, and renderer
npm run build      # production build

Native modules (SQLite, vector store) are compiled against Electron's ABI — run npm run rebuild after adding one.

Copy .env.example to .env for cloud inference, voice, and the integrations. It runs without any of them; those paths degrade gracefully rather than crashing.

Where this repo stands

This is the open base — the architecture, the memory tiers, the trigger engine, the desktop plumbing. The current polished build lives elsewhere while it's still moving fast, but everything structural is here.

Fork it, break it, rip out the memory layer and put something better in. If you build something interesting on top, I'd genuinely like to see it — open an issue and tell me.

Why I built it

Every assistant I used had the same failure: it was excellent once I knew what to ask, and useless in the ten minutes before that. The value was locked behind me noticing I needed help. I wanted to know whether you could move it earlier.

Three decisions came out of that:

Memory is a tiered hierarchy, not one embedding table. Similarity answers what is this like. A coworker needs what happened, when, and what I did about it — so recall walks tiers in cost order and stops at the first one that can actually answer.

The trigger engine's real job is staying quiet. Anything that interrupts you while you are concentrating has already lost, no matter how good the suggestion was. Most of that module is the decision not to speak.

Local inference comes first. The screen contents are the input here. Sending them to a cloud model by default was never an acceptable design, so the local path is the default and the cloud is the fallback.

License

MIT — see LICENSE. Do what you like with it.

_{Built by Rohit Maruri  ·  rohitmaruriats@gmail.com}