Hermes

Hermes has two ways to talk to Cortex — a native memory-provider plugin (recommended: Hermes calls into it automatically every turn) and a generic MCP tool registration (Cortex is one tool among others the agent chooses to call). Pick one per Hermes profile — running both against the same profile risks writing every turn twice.

Native memory provider (recommended)

The cortexlayer-hermes-plugin repo implements Hermes' MemoryProvider interface directly: memory is recalled before each LLM call and saved after, with no tool call for the agent to remember to make.

hermes plugins install cortexlayer
hermes plugins enable cortexlayer

install pulls in the plugin's Python dependency on its own — no separate pip install step.

Then run hermes memory setup, select cortexlayer, and paste an API key created on the Setup or Keys page of the Cortex web app. That's the only configuration — the key is written to ~/.hermes/.env as CORTEX_API_KEY and is what identifies your account; there's nothing else to map or configure.

This is Platform mode only for now: it talks to the hosted api.cortexlayer.net, not a self-hosted server, and there's no local/offline mode.

What Hermes gets

The same five tools as the MCP route below (memory_add, memory_retrieve, memory_update, memory_delete, memory_relink), available for Hermes to call explicitly — plus two things the MCP route can't do: automatic recall before every turn (no tool call needed) and automatic saving after every turn, including a checkpoint pass before Hermes compresses/truncates old context, so older turns aren't lost to compaction the way they would be otherwise.

MCP tool registration (alternative)

Hermes can also talk to Cortex over MCP like Claude Code, Cursor or any other MCP client — Cortex shows up as a set of tools the agent decides to call, rather than something Hermes calls into automatically. Create an API key the same way, then:

CORTEX_API_KEY=YOUR_API_KEY

mcp_servers:
  cortex:
    url: https://mcp.cortexlayer.net/mcp
    headers:
      Authorization: "Bearer ${CORTEX_API_KEY}"

Check the connection:

hermes mcp test cortex

Self-hosting Cortex instead of using the hosted server? Swap mcp.cortexlayer.net for your own server's URL — this route (unlike the native plugin above) works against a self-hosted server today.

Try it

1. Tell Hermes: "Use Cortex to remember that I prefer dark mode."
2. In a new conversation, ask: "What do I prefer for my editor theme?" (with the native plugin, you don't even need to ask it to check memory first — it already will have.)
3. Open Memory or Graph in the web app to see what was stored and how memories link together.