feat: replace Voyage AI with local ONNX embeddings (@xenova/transformers)

.env.example

@@ -1,5 +1,5 @@
 # =============================================================================
-# Harper Cortex - Environment Variables
+# Cortex - Environment Variables
 # =============================================================================
 # Copy this file to .env and fill in your values:
 #   cp .env.example .env
@@ -15,9 +15,9 @@ CLI_TARGET_PASSWORD=your_admin_password
 # Get your API key at https://console.anthropic.com/
 ANTHROPIC_API_KEY=sk-ant-xxxxx
 
-# --- Voyage AI (for vector embeddings) ---
-# Get your API key at https://dash.voyageai.com/
-VOYAGE_API_KEY=pa-xxxxx
+# --- Embeddings ---
+# Embeddings are generated locally using ONNX (Xenova/all-MiniLM-L6-v2).
+# No API key needed — runs in-process on first call (downloads ~23MB model once).
 
 # --- Slack (for webhook ingestion) ---
 # Create a Slack app at https://api.slack.com/apps

CLAUDE.md

@@ -1,4 +1,4 @@
-# Harper-Cortex
+# Cortex
 
 An agent-agnostic AI memory system using Harper Fabric as the vector database and MCP for AI agent connectivity.
 
@@ -7,7 +7,7 @@ An agent-agnostic AI memory system using Harper Fabric as the vector database an
 - **Runtime**: Harper (Node.js-based, ES modules)
 - **Database**: Harper Fabric with HNSW vector indexing
 - **Classification**: Anthropic Claude Haiku 3.5 (`claude-haiku-3-5-20241022`)
-- **Embeddings**: Voyage AI `voyage-3` (1024-dim)
+- **Embeddings**: `@xenova/transformers` with `all-MiniLM-L6-v2` (384-dim, Harper-native ONNX)
 - **Ingestion**: Slack Events API webhooks
 - **AI Bridge**: Harper MCP Server + `mcp-remote`
 - **Tests**: Node.js built-in test runner (`node:test`)
@@ -30,7 +30,7 @@ npm run deploy # Deploy to Harper Fabric
 
 ## Architecture
 
-Slack webhook -> classify (Claude) + embed (Voyage AI) -> store in Memory table -> query via MCP from Claude Desktop/Cursor.
+Slack webhook -> classify (Claude) + embed (@xenova/transformers) -> store in Memory table -> query via MCP from Claude Desktop/Cursor.
 
 ## Synapse
 
@@ -50,7 +50,7 @@ Universal Context Broker — ingests development context from any AI tool (Claud
 
 ### Conventions
 
-- SynapseEntry table follows same patterns as Memory (HNSW vector index, classification via Claude Haiku, embeddings via Voyage AI)
+- SynapseEntry table follows same patterns as Memory (HNSW vector index, classification via Claude Haiku, embeddings via @xenova/transformers)
 - Use renamed import: `const { SynapseEntry: SynapseEntryBase } = tables;`
 - All Synapse queries must filter on `projectId`
 - Default status filter is `active` (excludes superseded/archived)

CONTRIBUTING.md

@@ -1,13 +1,13 @@
-# Contributing to Harper-Cortex
+# Contributing to Cortex
 
 Thanks for your interest in contributing. This guide covers how to get set up, make changes, and submit a pull request.
 
 ## Getting Started
 
 1. Fork the repository and clone your fork:
    ```bash
-   git clone https://github.com/YOUR_USERNAME/Harper-Cortex.git
-   cd Harper-Cortex
+   git clone https://github.com/YOUR_USERNAME/Cortex.git
+   cd Cortex
    ```
 
 2. Install dependencies:

README.md

@@ -1,4 +1,4 @@
-# Harper-Cortex
+# Cortex
 
 A persistent, agent-agnostic AI memory system powered by [Harper Fabric](https://harper.fast). Clone, configure your API keys, deploy, and give all your AI tools a shared brain.
 
@@ -62,8 +62,8 @@ Create your free cluster at [fabric.harper.fast](https://fabric.harper.fast):
 ### 2. Clone and install
 
 ```bash
-git clone https://github.com/HarperFast/Harper-Cortex.git
-cd Harper-Cortex
+git clone https://github.com/HarperFast/Cortex.git
+cd Cortex
 npm install -g harperdb   # Install the Harper runtime (one-time)
 npm install               # Install project dependencies
 ```
@@ -75,7 +75,6 @@ Sign up for these services and grab your API keys. All have free tiers.
 | Service       | Sign Up                                                 | What You Need                                                                                        |
 | ------------- | ------------------------------------------------------- | ---------------------------------------------------------------------------------------------------- |
 | **Anthropic** | [console.anthropic.com](https://console.anthropic.com/) | API key (used for message classification)                                                            |
-| **Voyage AI** | [dash.voyageai.com](https://dash.voyageai.com/)         | API key (used for vector embeddings)                                                                 |
 | **Slack**     | [api.slack.com/apps](https://api.slack.com/apps)        | Create a Slack app. See [docs/slack-app-setup.md](docs/slack-app-setup.md) for the full walkthrough. |
 
 ### 4. Configure environment
@@ -123,13 +122,12 @@ See [docs/mcp-setup.md](docs/mcp-setup.md) for configuration instructions.
 | Variable               | Required        | Description                                                              |
 | ---------------------- | --------------- | ------------------------------------------------------------------------ |
 | `ANTHROPIC_API_KEY`    | Yes             | Anthropic API key for Claude (message classification)                    |
-| `VOYAGE_API_KEY`       | Yes             | Voyage AI API key (vector embedding generation)                          |
 | `SLACK_SIGNING_SECRET` | For Slack       | Slack app signing secret (webhook verification)                          |
 | `SLACK_BOT_TOKEN`      | For Slack       | Slack bot user OAuth token (`xoxb-...`)                                  |
 | `CLI_TARGET`           | For deploy      | Harper Fabric cluster URL (e.g., `https://cluster.org.harperfabric.com`) |
 | `CLI_TARGET_USERNAME`  | For deploy      | Harper cluster admin username                                            |
 | `CLI_TARGET_PASSWORD`  | For deploy      | Harper cluster admin password                                            |
-| `SYNAPSE_ENDPOINT`     | For Synapse CLI | Base URL of Harper-Cortex deployment                                     |
+| `SYNAPSE_ENDPOINT`     | For Synapse CLI | Base URL of Cortex deployment                                            |
 | `SYNAPSE_PROJECT`      | For Synapse CLI | Project ID to scope context entries                                      |
 | `SYNAPSE_AUTH`         | For Synapse CLI | Authorization header (e.g. `Basic dXNlcjpwYXNz`)                         |
 
@@ -207,21 +205,21 @@ Tests use Node.js built-in test runner with module mocking. No extra test depend
 
 1. **A source sends an event** via webhook (e.g. Slack message, GitHub issue, Linear task)
 2. **Classification**: Claude Haiku categorizes the content (decision, action_item, knowledge, etc.) and extracts entities (people, projects, technologies)
-3. **Embedding**: Voyage AI generates a 1024-dimensional vector embedding
+3. **Embedding**: A local ONNX model (`all-MiniLM-L6-v2`) generates a 384-dimensional vector embedding — no API key required
 4. **Storage**: Raw text, classification, entities, and embedding are stored in the Memory table with HNSW vector indexing
 5. **Retrieval**: Any MCP-connected AI client queries the Memory table using hybrid search (vector similarity + attribute filters)
 
 ### Synapse: coding context (CLI / API)
 
 1. **Ingest**: `synapse sync` reads your tool context files (CLAUDE.md, `.cursor/rules/`, `.windsurf/rules/`, `copilot-instructions.md`) and POSTs them to `/SynapseIngest`
 2. **Parse**: Each source format is split into discrete entries; duplicate content is deduplicated via content hash
-3. **Classify + embed**: Each entry is classified into a type (`intent`, `constraint`, `artifact`, `history`) and embedded with Voyage AI
+3. **Classify + embed**: Each entry is classified into a type (`intent`, `constraint`, `artifact`, `history`) and embedded locally with `all-MiniLM-L6-v2`
 4. **Storage**: Entries are stored in the SynapseEntry table with HNSW vector indexing, scoped by `projectId`
 5. **Retrieval**: `synapse search` or any MCP client queries `/SynapseSearch`; `synapse emit` formats entries back into any target tool's native format
 
 ## Supported Integrations
 
-This repo ships with Slack + Anthropic + Voyage AI as the default stack. The architecture is designed to be swappable - add a new webhook resource class for any ingestion source, or change the LLM/embedding provider in `resources.js`.
+This repo ships with Slack + Anthropic + local ONNX embeddings as the default stack. The architecture is designed to be swappable — add a new webhook resource class for any ingestion source, or change the LLM/embedding provider in `resources.js`.
 
 ### Ingestion Sources
 
@@ -254,12 +252,13 @@ Swap the classification model by changing `CLASSIFICATION_MODEL` in `resources.j
 
 Swap the embedding provider by changing `generateEmbedding()` in `resources.js`. If you change the vector dimensions, re-embed all existing records.
 
-| Provider           | Recommended Model      | Dimensions | Trade-off                       |
-| ------------------ | ---------------------- | ---------- | ------------------------------- |
-| **Voyage AI**      | voyage-3               | 1024       | Anthropic-recommended (default) |
-| **OpenAI**         | text-embedding-3-small | 1536       | Most widely adopted             |
-| **Cohere**         | embed-v4               | 1024       | Strong multilingual support     |
-| **Ollama** (local) | nomic-embed-text       | 768        | Full privacy, zero API cost     |
+| Provider                 | Recommended Model      | Dimensions | Trade-off                        |
+| ------------------------ | ---------------------- | ---------- | -------------------------------- |
+| **@xenova/transformers** | all-MiniLM-L6-v2       | 384        | Local ONNX, no API key (default) |
+| **Voyage AI**            | voyage-3               | 1024       | High quality, requires API key   |
+| **OpenAI**               | text-embedding-3-small | 1536       | Most widely adopted              |
+| **Cohere**               | embed-v4               | 1024       | Strong multilingual support      |
+| **Ollama** (local)       | nomic-embed-text       | 768        | Full privacy, zero API cost      |
 
 ### MCP Clients (Retrieval)
 
@@ -275,7 +274,7 @@ Any MCP-compliant AI client can connect to the Harper MCP Server and query your
 
 ## Synapse: Universal Context Broker
 
-Synapse extends Harper-Cortex into a **Universal Context Broker** — a system that bridges context across AI development tools. When you switch from Claude Code to Cursor, or add a new team member, the "Why" behind architectural decisions is normally lost. Synapse captures it.
+Synapse extends Cortex into a **Universal Context Broker** — a system that bridges context across AI development tools. When you switch from Claude Code to Cursor, or add a new team member, the "Why" behind architectural decisions is normally lost. Synapse captures it.
 
 ```
   INGEST (Tool → Harper)              EMIT (Harper → Tool)
@@ -311,7 +310,7 @@ Synapse extends Harper-Cortex into a **Universal Context Broker** — a system t
 # Install globally after cloning
 npm install -g .
 
-# Sync your context files to Harper-Cortex
+# Sync your context files to Cortex
 SYNAPSE_PROJECT=my-app synapse sync
 
 # Search across all context

bin/synapse.js

@@ -1,6 +1,6 @@
 #!/usr/bin/env node
 /**
- * Synapse CLI — Universal Context Broker for Harper-Cortex
+ * Synapse CLI — Universal Context Broker for Cortex
  *
  * Commands:
  *   synapse sync              Discover and ingest context files
@@ -10,7 +10,7 @@
  *   synapse status            Show entry counts by type and source
  *
  * Environment:
- *   SYNAPSE_ENDPOINT   Base URL of your Harper-Cortex deployment
+ *   SYNAPSE_ENDPOINT   Base URL of your Cortex deployment
  *   SYNAPSE_PROJECT    Project ID to scope entries
  *   SYNAPSE_AUTH       Authorization header value (e.g. "Basic dXNlcjpwYXNz")
  */
@@ -337,7 +337,7 @@ async function cmdWatch() {
 
 function printHelp() {
 	console.log(`
-Synapse — Universal Context Broker for Harper-Cortex
+Synapse — Universal Context Broker for Cortex
 
 Usage:
   synapse <command> [options]
@@ -359,7 +359,7 @@ Search options:
   --type <type>          Filter by type: intent, constraint, artifact, history
 
 Environment variables:
-  SYNAPSE_ENDPOINT       Harper-Cortex base URL (default: http://localhost:9926)
+  SYNAPSE_ENDPOINT       Cortex base URL (default: http://localhost:9926)
   SYNAPSE_PROJECT        Project ID (required for all commands)
   SYNAPSE_AUTH           Authorization header value (e.g. "Basic dXNlcjpwYXNz")
 

docs/architecture.md

@@ -18,7 +18,7 @@
 │  Linear         │ ──▶ │     │ classifyMsg (Claude API) │    │
 │  Webhooks       │     │     └────────────┬────────────┘    │
 └─────────────────┘     │     ┌────────────▼────────────┐    │
-                        │     │ genEmbedding (Voyage AI) │    │
+                        │     │ genEmbedding (ONNX local) │    │
 ┌─────────────────┐     │     └────────────┬────────────┘    │
 │  Other Sources  │ ──▶ │     ┌────────────▼────────────┐    │
 │  (Discord, etc) │     │     │      Memory.put()        │    │
@@ -61,7 +61,7 @@
 | `authorName`     | String     | No          | Human-readable author name                    |
 | `classification` | String     | Yes         | LLM category (decision, action_item, etc.)    |
 | `entities`       | Any (JSON) | No          | Extracted entities (people, projects, tech)   |
-| `embedding`      | [Float]    | HNSW/cosine | 1024-dim vector from Voyage AI                |
+| `embedding`      | [Float]    | HNSW/cosine | 384-dim vector from all-MiniLM-L6-v2 (local)  |
 | `summary`        | String     | No          | LLM-generated one-line summary                |
 | `timestamp`      | Date       | Yes         | When the original message was sent            |
 | `threadTs`       | String     | No          | Slack thread timestamp for grouping           |
@@ -101,7 +101,7 @@ Synapse adds a second data layer alongside Memory — the `SynapseEntry` table 
 │ .cursor/rules/   │     │  SynapseIngest                       │ ──▶ .cursor/*.mdc
 │ .windsurf/rules/ │     │  1. Parse (tool-specific parser)     │ ──▶ .windsurf/*.md
 │ copilot-inst.    │     │  2. Classify (Claude Haiku)          │ ──▶ copilot-inst.
-│ manual / slack   │     │  3. Embed (Voyage AI)                │
+│ manual / slack   │     │  3. Embed (ONNX local)               │
 └──────────────────┘     │  4. Store → SynapseEntry             │
                          └──────────────────┬───────────────────┘
                                             │
@@ -135,7 +135,7 @@ Synapse adds a second data layer alongside Memory — the `SynapseEntry` table 
 | `content`      | String   | No          | Full context text                                                 |
 | `source`       | String   | Yes         | `claude_code \| cursor \| windsurf \| copilot \| manual \| slack` |
 | `sourceFormat` | String   | No          | `markdown \| mdc \| json`                                         |
-| `embedding`    | [Float]  | HNSW/cosine | 1024-dim vector                                                   |
+| `embedding`    | [Float]  | HNSW/cosine | 384-dim vector from all-MiniLM-L6-v2 (local)                      |
 | `summary`      | String   | No          | LLM one-liner                                                     |
 | `status`       | String   | Yes         | `active \| superseded \| archived`                                |
 | `references`   | [String] | No          | Memory record IDs this traces back to                             |

docs/mcp-setup.md

@@ -43,7 +43,7 @@ Edit the Claude Desktop configuration file:
 ```json
 {
 	"mcpServers": {
-		"harper-cortex": {
+		"cortex": {
 			"command": "npx",
 			"args": [
 				"-y",
@@ -72,7 +72,7 @@ This returns `{ "jwt": "...", "refreshToken": "..." }`. Use the JWT in your conf
 ```json
 {
 	"mcpServers": {
-		"harper-cortex": {
+		"cortex": {
 			"command": "npx",
 			"args": [
 				"-y",

docs/slack-app-setup.md

@@ -1,12 +1,12 @@
 # Slack App Setup Guide
 
-This guide walks you through creating and configuring a Slack app to send messages to Harper-Cortex.
+This guide walks you through creating and configuring a Slack app to send messages to Cortex.
 
 ## Step 1: Create the Slack App
 
 1. Go to [api.slack.com/apps](https://api.slack.com/apps)
 2. Click **Create New App** > **From scratch**
-3. Name it (e.g., "Harper-Cortex Bot") and select your workspace
+3. Name it (e.g., "Cortex Bot") and select your workspace
 4. Click **Create App**
 
 ## Step 2: Configure Bot Token Scopes
@@ -60,7 +60,7 @@ If Slack asks you to reinstall the app after changing scopes or events, do so.
 In each Slack channel you want monitored, type:
 
 ```
-/invite @Harper-Cortex Bot
+/invite @Cortex Bot
 ```
 
 (Use whatever name you gave your app in Step 1.)

docs/synapse-design.md

@@ -1,10 +1,10 @@
-# Synapse: Universal Context Broker for Harper-Cortex
+# Synapse: Universal Context Broker for Cortex
 
 ## Context
 
 AI development tools (Claude Code, Cursor, Windsurf, Copilot) each build their own "contextual debt" — session history, rules files, memory banks — that is invisible to other tools. When engineers switch harnesses, the "Why" behind architectural decisions is lost.
 
-Synapse transforms Harper-Cortex into a **Universal Context Broker**: any tool can write structured context to Harper, and any other tool can read it in its native format. Harper becomes the single source of truth for development context.
+Synapse transforms Cortex into a **Universal Context Broker**: any tool can write structured context to Harper, and any other tool can read it in its native format. Harper becomes the single source of truth for development context.
 
 ## Architecture Overview
 
@@ -106,7 +106,7 @@ Accepts `{ source, content, projectId, parentId?, references? }`:
 
 1. Validates input
 2. Calls source-specific **parser** to split content into entries
-3. For each entry: classify (Claude Haiku) + embed (Voyage AI) in parallel
+3. For each entry: classify (Claude Haiku) + embed (all-MiniLM-L6-v2 via @xenova/transformers) in parallel
 4. Stores as SynapseEntry records
 
 **Parsers** (one per tool):
@@ -222,4 +222,4 @@ After each phase:
 - `npm test` — all existing + new tests pass
 - `npm run dev` — endpoints respond at `localhost:9926`
 - Manual: POST to `/SynapseIngest` with a CLAUDE.md, then `/SynapseSearch` to retrieve, then `/SynapseEmit --target cursor` to format
-- After Phase 4: `synapse sync && synapse search "architecture" && synapse emit --target cursor` in the Harper-Cortex project itself
+- After Phase 4: `synapse sync && synapse search "architecture" && synapse emit --target cursor` in the Cortex project itself