Aller au contenu

FAQ

Ce contenu n’est pas encore disponible dans votre langue.


Cortex is an open-source agentic memory and governance engine for LLM-driven teams. It stores, versions and serves structured knowledge (profiles, skills, methods, governance rules, memory) that LLM agents consume to work with continuity and accountability. See 01-why-cortex for the full value proposition.

Yes — under Apache 2.0 license. The code is available on GitHub. See LICENSE and NOTICE.

No. RAG (retrieval-augmented generation) retrieves relevant text chunks from a vector store and injects them into the LLM prompt. Cortex does something different: it stores typed, versioned, structured artefacts (skills, profiles, governance rules, memory entries) in a graph, not raw text chunks in a vector space. Cortex supports embeddings for semantic search, but its primary value is in the structured graph model, taxonomy and versioning — not in vector similarity.

You can use Cortex alongside a RAG pipeline. They serve different purposes.

No. Cortex is not an LLM. It is a knowledge and governance layer that LLM agents read from and write to. Your agents still call Claude, GPT, Gemini or any other provider for inference. Cortex adds structure, memory and rules to what those agents consume and produce.

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 instance.


Docker and Docker Compose (v2). At least 4 CPU cores and 8 GB RAM. No cloud account is needed. See 02-getting-started.

The services (API, MCP server, UI) are Node.js applications that can run natively. However, you would need to set up SurrealDB, PostgreSQL, MeiliSearch, MinIO and Ollama yourself. Docker Compose is the recommended approach.

It depends on the size of your catalogue, memory and document library. A minimal installation (empty catalogue, no embeddings) uses about 2 GB for Docker images and volumes. A production deployment with embeddings and thousands of artefacts may use 20-50 GB.

Can I run Cortex on ARM (Apple Silicon, Raspberry Pi)?

Section titled “Can I run Cortex on ARM (Apple Silicon, Raspberry Pi)?”

Yes. All Docker images used (SurrealDB, PostgreSQL, MeiliSearch, MinIO, Ollama, Node.js) publish ARM64 variants.


Cortex supports workspace and project isolation within a single instance. Each project has its own memory scope. However, the current architecture is designed for single-team use. There is no per-tenant database isolation or billing separation. For multi-organization deployments, run separate Cortex instances.

Yes. Each instance is independent with its own SurrealDB, PostgreSQL and other services. The divergence and cross-pollination features are designed precisely for this: comparing and merging catalogues across independent instances.

There is no hard limit. SurrealDB handles millions of records. The in-memory AtomIndex may become a factor at very large scales (hundreds of thousands of atoms). For most teams, this is not a concern.


No. No telemetry, no analytics, no machine identifier, no usage tracking. See 07-security-and-data.

Yes. Cortex has no mandatory external dependency. All services (including Ollama for embeddings) run locally. The only external calls happen when you configure a GenAI provider and a consuming service uses the decrypted key to call that provider’s API. If you do not configure providers, Cortex makes zero outbound network calls.

Provider API keys are encrypted with AES-256-GCM at rest. The encryption key is an environment variable you control. Keys are never returned in cleartext via the API or UI. See 05-providers for details.

Only what you explicitly configure. If you set up a GenAI provider, the consuming application sends the decrypted key and its payloads to that provider. Cortex itself sends nothing. See 07-security-and-data.


For catalogue reads (get atom, search, list blocs), expect 5-50ms on a local deployment. Memory reads are similarly fast. Ingestion (write path) takes 20-200ms per artefact depending on whether embedding is enabled.

Does embedding generation slow things down?

Section titled “Does embedding generation slow things down?”

Embedding is non-blocking and best-effort. If Ollama is slow or unavailable, ingestion completes without the embedding. The artefact is stored and indexed; the embedding is added asynchronously when available.

Yes. Remove the Ollama service from your Docker Compose stack. Cortex works without embeddings — you lose semantic search but retain all other features (full-text search, graph navigation, taxonomy).


Back up three components: SurrealDB (HTTP export endpoint), PostgreSQL (pg_dump), and MinIO (bucket mirror). See the backup section in 08-deployment.

For production, at least twice daily. For active development, once daily is usually sufficient.

Yes. Import the SurrealDB dump via the HTTP import endpoint, restore PostgreSQL with psql, and sync the MinIO bucket back. Stop the API and MCP services during restore, then restart.


See CONTRIBUTING.md. Commit conventions, DCO sign-off (git commit -s), PR workflow.

Dedicated procedure in SECURITY.md. Do not open a public issue for a vulnerability.