Introduction

SubCortex is cognitive infrastructure for AI agents and the people who use them. It provides persistent memory, architectural enforcement, identity management, and relationship tracking — a brain that travels with your agents and your people across tools, sessions, and teams.

What it solves

AI agents are amnesiacs. Every session starts cold. Every tool is a silo. The agent you spent an hour teaching your architecture to? It forgot everything when you closed the tab.

Static instruction files (CLAUDE.md, .cursorrules) are a band-aid — they don't evolve, they don't track decisions, and the model might ignore them. Your team's architectural rules are advisory, not enforced. Institutional knowledge lives in senior engineers' heads and walks out when they leave.

SubCortex solves this at three levels:

Personal Brain

A portable cognitive profile keyed to your identity. Your technology preferences, working style, expertise areas, and decisions — carried across every tool that supports SubCortex. Claude Code today. Cursor tomorrow. Any MCP-compatible agent. The brain gets smarter the more you use it, and you own the data.

Project Brain

Architectural rules, enforcement gates, and decision history scoped to a codebase. "All API routes must call withAuth()" isn't a suggestion the model might follow — it's a structural constraint that blocks the tool call before bad code is written. Decisions are tracked with provenance, confidence scores, and temporal context so your team can query "what did we decide about auth last month, and why?"

Enterprise Brain

The full cognitive layer for teams building AI-powered products. User rapport, relationship graphs, emotional intelligence, compliance-grade audit trails. Tenant-isolated, SOC2-ready, with the same assertion engine underneath.

Core Primitives

Assertions — facts about the world. A user's role, a project's architectural constraint, a developer's preference. Stored with confidence scores, temporal bounds, and provenance tracking. Immutable — changes create new assertions, never overwrite.

Rules — typed code guardrails stored as assertions. Banned patterns, required patterns, file constraints, propagation requirements. Enforced structurally at the tool level, not advisorily in the prompt.

Relationships — connections between subjects. Who reports to whom, which module depends on which, what entities are linked. A living knowledge graph the agent reasons over.

Signals — emotional and behavioral state with decay. Not just what happened, but how the user feels about it right now. Signals decay over time toward neutral.

Thalamus — the cognitive intake pipeline. Raw information from agents is classified, deduplicated, checked for contradictions, and stored intelligently. Conflicts are surfaced, not silently overwritten.

Architecture

Your Agent / IDE / Tool
    ↓
SubCortex SDK (@subcortex-ai/sdk)
    ↓
SubCortex Engine (REST + gRPC)
    ↓
Assertion Store + Signal Index + Mesh Graph + Rules Engine

The engine exposes REST APIs. The SDK handles auth, retries, and context generation. Plugins and hooks wire it into your tools. You never talk to the engine directly.

Next steps

• Quick Start — running in 5 minutes
• Authentication — API keys and tenant context
• Core Concepts — assertions, signals, relationships in depth
• The Portable Agentic Brain — the personal cognitive layer