Cortex

What Cortex MCP server does

The Cortex MCP server gives AI assistants a persistent memory space that can be shared across later conversations and different compatible clients. It is designed to retain decisions, context, and source references, then retrieve them when a later question depends on previous work.

Rather than treating memory as an unstructured log, the hosted service checks new entries for novelty and redundancy, extracts typed claims, and tracks contradictions between claims. Memories are connected through entities and typed relationships. Each memory also receives a coherence score describing how well it fits the existing knowledge base. A background consolidation process turns episodic memories into longer-term knowledge.

The memory space is personal to each user. The provider states that memories are isolated between users, can be exported or deleted on request, and are stored on infrastructure operating in the European Union.

How it works

The Cortex MCP server is available as a remote streamable HTTP endpoint. Claude users can add the endpoint as a custom connector and authenticate through Google or email; the account and personal memory space are created during first sign-in. Claude Code can connect with its HTTP MCP configuration.

Clients without native remote connector support can run mcp-remote as a stdio bridge. The repository also includes a local Node.js stdio process that forwards tool calls to the hosted service. That process can start and list its tools without credentials, but forwarded calls require an access token.

The service queues memory writes when needed. The write-status tool reports the quality-gate result and any unresolved conflicts. Deletion is exposed separately and requires explicit confirmation.

Setup and configuration

For a direct Claude connection, add this endpoint as a custom connector:

https://skynetlab-cortex-saas-mcp.cortex-320.workers.dev/mcp

For Claude Code, use the HTTP transport command shown in the project documentation. A generic client can configure mcp-remote with the same URL. To run the local bridge from source, install the Node.js dependencies and start server/index.js with CORTEX_ACCESS_TOKEN set. Docker is also supported.

CORTEX_ACCESS_TOKEN contains the OAuth bearer token issued at sign-in. It is optional for starting the local process, but required for authenticated forwarded calls. CORTEX_MCP_URL changes the remote destination and otherwise defaults to the hosted Cortex endpoint. Running npm test performs a basic initialization, tool-listing, and unauthenticated-call check.

Tools and capabilities

The seven exposed tools are:

• Search memories using semantic matching.
• Fetch one memory with its associated sources.
• Recall and synthesize what is known about a topic.
• Show contradictions tracked between memories.
• Save a memory after duplicate and redundancy checks.
• Check the result of a queued write.
• Forget a memory after explicit confirmation.

These operations make the Cortex MCP server suitable for assistants that need traceable recall, decision history, and visibility into conflicting information.

Limitations and notes

The memory engine is hosted and its source code is not published. The repository contains documentation, configuration examples, and the local stdio bridge rather than the engine itself. The hosted service requires authentication for normal memory operations, and removing a connector does not automatically delete the associated memories; deletion must be requested from the provider. The project documentation describes the engine as patent-pending and identifies the repository contents as MIT-licensed.