Cortex

Every time you start a new AI chat, you re-explain everything. Cortex fixes that.

It runs silently in the background, captures what you work on, and automatically surfaces relevant context the next time you prompt your AI assistant with a related topic. No copy-pasting, no manual notes, no setup per project.

How it works

Manual capture: Select any text from a file and press Cmd+Shift+C to save it as a memory. All memories appear in the Cortex sidebar in your activity bar.

Local embeddings: Each capture is embedded using all-MiniLM-L6-v2 (via @xenova/transformers) and stored in a local SQLite database. Nothing leaves your machine.

MCP server: Cortex exposes a search_memories tool and a save_memory tool over the Model Context Protocol. Any MCP-compatible client can call them.

Automatic context injection: The repo ships with rules that tell the AI what to do automatically. No prompt engineering required on your end.

At the start of every conversation, the AI calls search_memories based on what you ask and anything relevant comes back as context before it responds. After every substantive response (a fix, implementation, diagnosis), it calls save_memory with a 2–3 sentence summary of what it did and why. Over time, context accumulates across sessions.

Cortex works in both Cursor (via .cursor/rules/cortex-memory.mdc) and Codex (via AGENTS.md). Both files ship with the repo and are already active when you clone it.

Setup

1. Install the extension

Download the latest .vsix from Releases, then in Cursor open the Command Palette and run Extensions: Install from VSIX.

2. Register the MCP server

Use the full path to node to avoid PATH issues (e.g. if you use nvm).

Cursor — ~/.cursor/mcp.json:

{
  "mcpServers": {
    "cortex": {
      "command": "/path/to/node",
      "args": ["/path/to/cortex/out/mcp-server.js"]
    }
  }
}

Run Cortex: Copy Cursor MCP Config from the Command Palette to copy a pre-filled snippet with the correct mcp-server.js path.

Codex — ~/.codex/config.toml:

[mcp_servers.cortex]
command = "/path/to/node"
args = ["/path/to/cortex/out/mcp-server.js"]

Find your node path with which node.

3. Optional — Ollama summarization

By default, Cortex embeds your raw captured text. With Ollama, it first summarizes the text into a concise description before embedding, which means the semantic search finds more relevant memories and filters out noise. Recommended if you're capturing long functions or dense code blocks where raw text would produce noisy embeddings.

Install Ollama and pull the default model:

brew install ollama
ollama pull llama3
ollama serve

To use a different model, set cortex.ollamaModel in VS Code settings (default: llama3). Set it to an empty string to disable summarization even if Ollama is running.

What it captures

• Manual selections via Cmd+Shift+C or Cortex: Capture Selection
• AI chat summaries written by the MCP save_memory tool after substantive responses

Managing memories

The Cortex sidebar (activity bar icon) shows all saved memories with their source and timestamp. Memories saved via MCP from AI chats appear automatically.

Available commands (open Command Palette with Cmd+Shift+P):

• Cortex: Search Memories — semantic search across all memories
• Cortex: List Memories — view all memories in the output panel
• Cortex: Delete Memory — select and delete a specific memory
• Cortex: Clear All Memories — delete all memories (requires confirmation)
• Cortex: Show Top Result for Current File — quick relevance check for the active file

Memories are automatically pruned after 30 days if they've never been accessed.

Built with

Fully local. No API keys. No telemetry.