Introduction to Cortex
The shared memory layer for AI agents — governed, cited, and auditable.
Cortex is a unified, real-time shared memory layer designed for collaborative engineering teams and autonomous AI agents. Instead of forcing agents to reconstruct codebase contexts from raw file stuffing, Cortex exposes a semantic vector repository containing your team's design specifications, API contracts, and engineering decisions.
Unlike code indexes that parse syntactic details (classes, variables), Cortex indexes semantic context — why a design was chosen, standard error formats, and API scopes.
Key Features
Cited Memory Recall
Every query returns memories with source URL, author, and timestamp. No hallucinations — every answer is traceable.
Role-Based Access Control
Scoped permissions by seniority, project, and agent identity. Denials are explicit and logged.
MCP Server
Native Model Context Protocol endpoints allowing Claude, Cursor, and other LLMs to fetch code contexts on demand.
Developer CLI
Read, write, list, and search memories directly from your shell inside CI/CD workflows.
Audit Logging
Every read, write, and denial logged with timestamp, role, agent, and result. Full provenance tracking.
Conflict Detection
Automatic detection of contradictions, duplicates, and stale memories with authority-based resolution.
How It Works
graph LR
A[AI Agent] -->|MCP| B[Cortex Gateway]
B -->|Policy Check| C[RBAC Engine]
C -->|Scoped| D[Vector Search]
D -->|Cited| E[Memory Results]
E -->|Logged| F[Audit Trail]
- Agent sends a query via MCP or REST API
- Policy engine resolves the agent's allowed scopes based on role and clearance
- Vector search finds semantically relevant memories within permitted scopes
- Citations are attached to every result (source, author, timestamp)
- Audit log records the entire interaction
Use Cases
| Use Case | Description |
|---|---|
| Codebase Onboarding | New engineers query architectural decisions with full context |
| AI-Assisted Development | Claude/Cursor fetch design specs before writing code |
| Knowledge Preservation | Institutional knowledge survives team changes |
| Compliance & Audit | Every memory access is traceable and attributed |
| Multi-Agent Coordination | Agents share context while respecting access boundaries |
Architecture Overview
Cortex uses a dual-storage architecture:
- PostgreSQL — Authoritative store for memories, policies, users, and audit logs
- Qdrant — Vector database for semantic search with permission-filtered queries
The Memory Orchestrator sits at the center, enforcing policies on every read and write operation. No code path bypasses the policy engine.