Cortex Runtime

Cortex Runtime is a modular AI execution engine designed to build intelligent assistants that can think, remember, and act.

It combines:

• LLM reasoning
• semantic memory (embeddings)
• pluggable modules (actions)

---

Features

• LLM-powered reasoning with structured outputs
• Semantic memory with embeddings
• Modular architecture (core + external plugins)
• CLI-based plugin system
• Action execution pipeline

---

Architecture

---

Response Format

Cortex uses structured JSON responses from the LLM:

{
  "response": string,
  "actions": [
    {
      "type": string,
      "args": { /* specified by module */ }
    }
  ],
  "memory_candidates": [
    {
      "summary": string,
      "importance": float
    }
  ]
}

---

Modules

Modules are responsible for executing actions.

Types

• Core modules - built into the runtime
• External modules - standalone CLI programs

---

External Module Protocol

Modules communicate with Cortex via CLI:

Describe module

module --describe

Returns:

{
  "name": string,
  "description": string,
  "keywords": [string],
  "args_schema": {}
}

---

Run module

module --run { "room": "1", "state": "on" }

Returns:

{
  "status": "ok"
}

---

Memory System

Cortex uses embeddings for semantic memory:

• stores user facts and preferences
• retrieves relevant context using similarity search
• injects memory into LLM prompts

---

Getting Started

git clone https://github.com/kirbodevv/cortex-runtime
cd cortex-runtime
cargo run

---

Roadmap

• Persistent memory (database)
• Better memory ranking
• Modules execution