Skip to main content
Both settings live under Advanced options in the create flow. Neither is required — an agent works without touching either. Your agent keeps memory files regardless of these settings. What this section controls is whether it can search that memory semantically (by meaning) or only by keyword. Semantic search needs an embedding model, which is a separate service from the chat model, billed separately.

​
Embedding provider

The picker offers OpenAI, Gemini, Voyage, Mistral and DeepInfra. If one of them is already your agent’s chat provider, it is listed first and marked no extra key needed — the credential your agent already holds is reused. Choosing any other provider adds a field for that provider’s API key. The key is stored alongside your chat credential without replacing it, so your chat model keeps working.
“No extra key needed” does not mean free. Embeddings are billed separately by the provider, on the same account as your chat model.
Providers ClawUp cannot supply a credential for — Amazon Bedrock (AWS credential chain), Ollama and LM Studio (a server running inside your container), GitHub Copilot (a subscription token) — are deliberately not listed. There would be no way to make them work from this screen.

​
Subscription agents

If your agent uses a ChatGPT or Claude subscription rather than an API key, there is no free embedding option, and the form says so:
  • A ChatGPT/Codex subscription does not cover embeddings. They are a separately billed API-platform endpoint, and the subscription carries no API credit.
  • Anthropic publishes no embeddings API at all.
Memory search still works on keywords. For semantic search, pick any provider above and supply its key.

​
Choosing a model

Each provider offers several models, and the default is marked. Larger models recall better and cost more — Qwen3-Embedding comes in 8B, 4B and 0.6B through DeepInfra if you want to make that trade-off explicitly.

​
Memory engine

The engine is where memory is stored. Built-in is the default and needs no setup. The alternatives are plugins. Selecting one installs it into your agent: Cloud services listed as (coming soon) — Zep, Pinecone, Ragie, LlamaCloud — are not selectable yet. Engines that need an account, such as Mem0 or Supermemory, take their API key in the fields that appear when you select them.

​
Identity

An identity gives a new agent a persona and a set of skills in one step. Pick one under Choose an Identity when creating an agent, or apply one later with Install Identity on the agent’s card.
  • The persona tells the agent who it is and how to behave. It is delivered as the agent’s SOUL.md, so the agent can describe its own role when asked.
  • The skills are installed into the agent’s runtime. You can untick any you do not want before creating.
Applying an identity to an existing agent replaces its persona and adds the identity’s skills.

​
Skills on Hermes agents

Identity skills come from ClawHub, and not all of them install on Hermes. Some names are ambiguous in the registries Hermes searches, and Hermes’ security scanner blocks skills it rates dangerous. Skills that fail are reported with the reason; the rest install normally, and the persona is applied either way. Hermes ships its own catalogue of official skills, which install cleanly. Browse it inside the agent with hermes skills browse.