How VEKTOR Notes works

VEKTOR Notes is a local-first AI note-taking app for Android. Everything — your notes, memories, and the AI models processing them — runs on your device. No cloud uploads. No tracking.

The interface has two modes you swipe between: JOT for capture and CHAT for retrieval. Under the hood, a four-layer memory graph (MAGMA) builds a persistent, searchable representation of everything you write.

First launch

Install from Google Play and open the app. You land directly on the JOT surface — a blank text area. No account required. No onboarding wizard. The app is immediately functional for capture and local storage without any configuration.

To enable AI features (ghost suggestions and CHAT responses), you need to add an API key from a supported LLM provider.

Setting your API key

VEKTOR Notes uses your own provider keys. Your key is stored in SecureStore on-device — it never touches VEKTOR servers. API calls go directly from your device to the provider.

Supported providers

JOT — Zen capture

JOT is the primary writing surface. A clean text area with no menus, no formatting toolbar, no friction. You type. The app watches quietly. After 900 milliseconds of silence — long enough to fire only when you have genuinely paused, short enough that the suggestion arrives before you lose the thread — the ghost suggestion engine offers a completion.

JOT actions

CHAT — Talk to your memory

Swipe left from JOT to enter CHAT mode. Ask questions. The app searches everything you have written using a dual-channel retrieval pipeline and answers from your own context — not the internet.

• “What did I think about the authentication bug last week?”
• “What connects my notes on sleep and my notes on decision-making?”
• “What was I working on before I made that architecture decision?”
• “Summarise everything I've written about LLM memory systems.”

GRAPH — See connections

Swipe up from JOT to open the memory graph. Every saved memory is a node. Every inferred relationship is an edge. Nodes are coloured by layer type.

The retrieval pipeline

When you ask something in CHAT, two parallel retrieval paths run before the LLM sees your question:

MAGMA — The memory graph

MAGMA is the four-layer graph architecture that separates VEKTOR Notes from a notes app with a chat window bolted on. Each layer represents a different type of relationship, which determines how memories are retrieved.

Storage

The entire graph lives in a single SQLite file on your device. No cloud database. No GPU. Queries execute in milliseconds and back up with your device backup.

notes        -- raw note content + timestamps
memories     -- MAGMA nodes (content, layer, importance)
edges        -- typed directional relationships
entities     -- extracted named entities

FTS5 index   -- BM25 full-text search
vec_memories -- float32 embeddings (sqlite-vec ANN)

MCP bridge (Pro)

The MCP bridge connects VEKTOR Notes to Claude Desktop, Cursor, or any MCP-compatible agent on your computer over your local Tailscale network. Your phone's memory becomes available to desktop AI tools.

VEKTOR_DB_PATH=/path/to/vektor.db node vektor-notes-mcp-bridge.js

LLM provider routing

The model picker in Settings lets you route tasks to different models. JOT ghost suggestions can run on a faster, cheaper model than CHAT synthesis — you are not paying GPT-4 rates for a 30-word autocomplete.

What stays on your device

• All notes and memories are stored in SQLite on-device only
• Your API key is stored in Android's encrypted SecureStore
• API calls go directly from your device to your LLM provider — not via VEKTOR servers
• VEKTOR does not log your conversations, queries, or note content
• No analytics SDK, no ad network
• The app works fully offline for capture and local search