Why Cortex
Cortex is an open-source agentic memory and governance engine for LLM-driven teams.
The problem
Section titled “The problem”LLM agents are powerful executors. They reason, write code, draft documents, audit spreadsheets. Yet the moment a session ends, the context evaporates.
- No persistent memory. The agent forgets what it decided yesterday.
- No structured identity. Each session starts with a hand-written system prompt instead of a versioned, reusable profile.
- No shared catalogue. Skills, methods and expertise are copy-pasted between prompts instead of stored, versioned and composed.
- No governance trail. When an agent makes a decision, there is no trace, no version, no way to audit what happened and why.
- No cross-session learning. Patterns observed by one agent are invisible to the next.
These gaps compound. As teams scale from one agent to five, from one project to ten, the cost of improvisation becomes unsustainable.
What Cortex adds
Section titled “What Cortex adds”Cortex is a server-side layer that stores, versions and serves everything your agents need to work with continuity and accountability.
Typed atomic catalogue
Section titled “Typed atomic catalogue”Skills, expertises, methods, tools and scopes are stored as atoms — typed, versioned, queryable artefacts. Each atom has a stable identifier, a semantic version, a content hash for delta detection, and a six-layer structure (identity, classification, shape, presence, lifecycle, enrichment). You do not paste prompt snippets; you compose reusable building blocks.
HYOS profiles
Section titled “HYOS profiles”An agent identity is not an ad-hoc system prompt. It is a HYOS profile — a structured entity that carries a mission, a mindset, scoped-in and scoped-out boundaries, responsibilities, and a composition of blocks. Profiles are versioned. They can be created from templates (modeles), customized (sur mesure), or derived (variantes).
Living memory
Section titled “Living memory”Cortex stores per-project memory: decisions, errors and their corrections, recurring patterns, blockers and workarounds, needs, feedbacks, and inter-agent contradictions. Memory entries are written by agents during sessions and read back at the start of the next one. Sessions do not start from zero.
Governance and versioning
Section titled “Governance and versioning”Every artefact has a semantic version (version_semver) and a content hash. Mutations bump versions automatically based on a change classifier (patch, minor, major). Snapshots capture the state of any catalogue subset. Divergence reports compare two snapshots across five axes (topology, catalogue, profiles, governance, learning). Cross-pollination signals detect when two independent instances converge on the same solution and recommend promoting it to the shared catalogue.
Two API surfaces
Section titled “Two API surfaces”Cortex exposes a REST API (port 9096, key-based authentication) for programmatic access and a Model Context Protocol server (port 3001, OAuth 2.0) for direct LLM client integration. Claude, ChatGPT or Cursor connect to the MCP server; CI pipelines and custom services call the REST API.
3D graph UI
Section titled “3D graph UI”A React + Vite web interface provides visual exploration of the catalogue, full taxonomy navigation, profile cards, memory views and workspace management.
What Cortex is not
Section titled “What Cortex is not”Understanding what Cortex does not do is as important as understanding what it does.
- Not a chatbot. Cortex does not chat with end users. It is a backend service that stores and serves structured knowledge to agents that do.
- Not a workflow engine. Cortex does not orchestrate task execution. Orchestration products like arka-deck consume Cortex as their knowledge and governance layer. Cortex stores the rules; the orchestrator enforces the sequence.
- Not a vector database. Cortex supports embeddings (via Ollama and bge-m3), but its primary model is a typed graph with explicit taxonomy, not a flat vector space. Atoms have identity, version, classification and lifecycle — not just coordinates.
- Not a prompt library. Profiles and blocks are versioned graph entities with stable identifiers, content hashes and taxonomy edges — not markdown files in a folder.
- Not a cloud-only product. Cortex runs entirely on your infrastructure. The default deployment is a Docker Compose stack on a single machine. No external account is needed.
Differentiation
Section titled “Differentiation”| Capability | Vector DBs (Pinecone, Weaviate, …) | Chatbot platforms (ChatGPT, Claude.ai) | Workflow engines (n8n, Temporal, …) | Cortex |
|---|---|---|---|---|
| Typed artefact graph | — | — | — | Yes |
| Per-artefact semver | — | — | — | Yes |
| Agent profiles (HYOS) | — | System prompt | — | Versioned entity |
| Living per-project memory | — | Chat history | — | Structured entries |
| Governance rules | — | — | — | Atom-level, auditable |
| Divergence + cross-pollination | — | — | — | Built-in |
| MCP server for LLM clients | — | Native chat | — | Yes |
| REST API for scripts/CI | — | API access | API access | Yes |
| Self-hosted by default | Managed cloud | Managed cloud | Self-hosted | Self-hosted |
Who uses Cortex
Section titled “Who uses Cortex”Cortex is the engine behind arka-deck (multi-agent orchestration console), Pack PRO (sector-specific expert catalogues) and Cortex Lite (local sidecar). It is released as open source so any team building agentic software can run their own catalogue and memory layer.
If you build LLM agents that need:
- a shared, typed memory across sessions;
- a catalogue of reusable cognitive blocks;
- profiles to describe who agents are and what they are allowed to do;
- versioning so you can roll back;
- governance so you can audit;
then Cortex is for you.
What next?
Section titled “What next?”- Install in ten minutes: 02-getting-started
- Master the vocabulary: 03-concepts
- Pick a use case: 04-use-cases