The Unified Intelligence Layer for the Agentic Enterprise.

Cortex is a real-time, shared global context engine designed to solve the "Context Silo" problem in corporate AI workflows. Built for soon hackathon #1, Cortex ensures that every agentic interaction—from engineering to finance—is powered by a single, evolving source of truth.

In the modern enterprise, context is managed poorly. Individuals manually feed agents files, rules, and snippets session-by-session. This leads to:

• Inconsistency: Different team members receive different answers for the same policy.
• Wasted Time: Context is re-uploaded and re-explained every single time a new chat starts.
• Lost Knowledge: Critical decisions made in one session vanish the moment the window is closed.

Cortex solves this. It turns individual interactions into institutional memory.

• Shared Global Memory: A real-time graph of company context that accelerates development across all departments.
• Departmental Awareness: Tailored context for Engineering, Management, Finance, Marketing, and more.
• Autonomous Context Extraction: Using claude-haiku-4.5, Cortex automatically identifies "substantial" info (decisions, budgets, policies) from chat logs and saves them to the bank.
• Peer Experience Injection: Naturally surfaces similar situations encountered by colleagues (e.g., "Interestingly, [Name] from [Dept] ran into something similar...").
• "GitHub for Context": Managers can review "sub-nodes" (session context) and merge them into "department-nodes" (mainline truth).
• Multimodal Pipeline: Full support for image analysis and server-triggered Cloudinary transformations (blur, background removal, etc.) before the LLM even sees them.

1. Identity: User signs in. The proxy.ts middleware validates the session and injects x-cortex-department and x-cortex-name headers into every request.
2. Hydration: As development starts, Cortex fetches relevant context from the Context Bank (Supabase) and existing threads.
3. Real-time Synthesis: The backboard.ts engine uses keyword-overlap similarity to find relevant peer experiences and merges them into the system prompt.
4. Autonomous Storage: If a session yields new important data, it is identified by Haiku and stored as a Sub-node.
5. Governance: Managers review Sub-nodes, merging them into the Department-node or deleting irrelevant noise.

Cortex organizes information hierarchically:

• Sub-node: Transient context from a specific session or individual development task.
• Department-node: The established "Main" branch for a specific department (e.g., Engineering standards).
• Context-Map: The global company graph connecting all department nodes into a unified corporate brain.

Cortex leverages the latest Next.js 16 features. We utilize a proxy.ts pattern to handle secure auth and department-based routing, ensuring the LLM always knows the specific departmental constraints of the user.

• Primary LLM: claude-sonnet-4-6 for high-reasoning chat and multimodal analysis.
• Extraction: claude-haiku-4-5-20251001 for low-latency JSON extraction of context from messages.
• Memory Management: Integrated with Backboard.io to provide a persistent, "Readonly" or "Auto-writing" memory stream that lives outside the chat window.

• Database: Supabase for structured context entry storage and agent metadata.
• Caching: Backboard's thread-based memory provides a high-speed cache for institutional knowledge.
• Secret Management: Built-in warmSecretCache in instrumentation.ts for enterprise-grade security handling.

• Node.js 20.9+
• Anthropic API Key
• Backboard API Key
• Supabase Account

npm install

Create a .env.local file:

ANTHROPIC_API_KEY=your_key
BACKBOARD_API_KEY=your_key
NEXT_PUBLIC_SUPABASE_URL=your_url
NEXT_PUBLIC_SUPABASE_ANON_KEY=your_key
SUPABASE_SERVICE_ROLE_KEY=your_key
CORTEX_MCP_TOKEN=your_internal_token

npm run dev

• Claude CLI & Agentic IDE Integration: Feed Cortex context directly into your terminal or VS Code.
• Slack Real-time Sync: Bidirectional context flow between Slack channels and the Context Bank.
• Automated Conflict Detection: Alerting managers when a new sub-node contradicts existing department-node policies.

Built with ❤️ for soon hackathon #1