MemVector Knowledge Engine

Map what you think connects — and why.

Obsidian's graph shows what you linked. MemVector shows what you believe: which ideas imply each other, which depend on each other, which contradict. Declare a relation, and your notes move on a 2D canvas — related ideas pull together, conflicting ones push apart. Your thinking becomes a map you can navigate, query, and ask an AI about.

• Relations you can read. Every relation is a plain Markdown note with a type and a reason.
• Meaning as layout. Typed relations plus semantic similarity drive the canvas.
• Ask your own map. AI synthesis uses your declared structure as context.
• Fully local. Embeddings and graph live in one SQLite file in your vault. Works with Ollama, no API key needed.

---

How it works

You are writing a thesis and have 200 notes, but no overview of which claims support each other and which collide. Select two notes and declare how they relate:

Utilitarianism        →  CONFLICTS_WITH  →  Kantian duty ethics
Natural selection     →  REQUIRES        →  Heritable variation
Bayes' theorem        →  GENERALIZES     →  Likelihood ratio

That one step does three things:

1. The canvas rearranges. EQUIVALENT_TO pulls notes close, CONFLICTS_WITH pushes them apart. The layout is your declared structure.
2. A Markdown note appears in your vault (wiki/relations/ by default), with WikiLinks to both notes and your reason in plain text — readable, editable, searchable, version-controlled.
3. The relation joins a local graph that multi-hop lookups and AI synthesis can traverse.

Underneath, semantic embeddings act as a second force: notes with similar content drift together even without a relation. Your declared relations win where they are stronger; similarity fills the gaps.

MemVector does not replace Obsidian's graph, and it does not build the graph for you — you decide what connects. The AI only steps in when you ask it to synthesize.

---

Installation & Quickstart

1. In Settings → Community plugins → Browse, search for MemVector Knowledge Engine, install and enable it.
2. Choose your embedding and LLM provider — Ollama with bge-m3 works out of the box with no API key.
3. Click "Index entire vault" to compute embeddings.
4. Open the 2D canvas from the ribbon icon or command palette.
5. Select two notes (Cmd-click or Shift-drag lasso), then open the Relation Builder to declare your first typed relation.

---

Relation types

Seventeen canonical types ship as a starting vocabulary. All are customizable via wiki/relation-types.json in your vault — nothing is hardcoded.

Each type has a dedicated color on the canvas. Custom types get a deterministic color from a hash — no manual color assignment needed.

Force and repulsion are per-type data, not code: edit the weight / repels columns in Settings → Relation types, or the vocabulary JSON directly. Add your own types there, or switch the whole vocabulary to a bundled preset (STEM, Law, Medicine, Philosophy) — existing edges keep working, since edge labels are stored as plain strings.

---

Key features

• Intentional 2D layout. A force simulation combines your typed relations and semantic embeddings. Position is meaning.
• Relations as Markdown. Every relation is a .md file in your vault with a type, an optional reason, and WikiLinks to both notes. Not a black box.
• Three topology modes. Hub (Focal → Rest), convergence (Rest → Focal), or sequence (Chain: A → B → C → ...) — for structuring how a set of selected notes relate to each other.
• Mini-Radar sidebar. A polar view centered on your active note. Radial distance = true cosine distance (1 − similarity). Updates as you switch notes.
• Hybrid GraphRAG synthesis. When you ask the AI to synthesize, it enriches your selection with two independent lookups: vector-similar notes (finds semantic neighbors with no graph path) and graph-hop neighbors (finds structurally linked notes with different vocabulary). Neither alone covers both.
• 100% local-first, zero setup. All embeddings and graph edges live in a bundled SQLite database (memvector-local.sqlite) inside your vault. No Docker, no server, no network call for storage.
• Flexible AI providers. Ollama (local), Anthropic Claude, OpenAI, DeepSeek, OpenRouter, or any custom OpenAI-compatible endpoint. API keys are stored in Obsidian's secure Secret Storage — never in plain text.

---

Status: Early-Stage (Pre-1.0)

This plugin is at version 0.3.x and under active development. The version number stays below 1.0.0 on purpose — it has not yet reached the stability that number implies. Expect rough edges, and check the open issues before relying on it for anything critical.

---

Documentation

• Roadmap (docs/ROADMAP.md) — v0.1 capabilities and future milestones
• Architecture (docs/ARCHITECTURE.md) — component map, force layout, SQLite storage
• Configuration (docs/CONFIGURATION.md) — settings reference, provider setup, vocabulary
• User Guide (docs/USER_GUIDE.md) — canvas controls, gesture reference, synthesis workflow
• Hybrid GraphRAG (docs/GRAPHRAG.md) — how synthesis enriches prompts with vector and graph context

Development & Contributing

• Runbook (docs/runbook.md) — clone, build and test commands, linking the clone into a vault, the local mock server, inspecting the index, releasing
• Manual testing (docs/TESTING.md) — smoke test steps inside Obsidian

---

License

Distributed under the MIT License.