Skip to content

Skills and memory

Skills describe how to carry out a task. Memory keeps the facts and reference notes that are useful across conversations. Each section opens on its own page; the breadcrumb takes you back to the list.

  1. Open Skills and search for its name or description.
  2. Select the skill to read its instructions.
  3. If it is editable, change its content, then choose Save.

Skills included in Hermes are read-only. To adapt one, ask your agent to create a custom skill. Editable persistent skills are saved in that agent’s capsule. If the content has changed in the meantime, the portal refuses to overwrite that version: keep your text, reload the skill and reconcile the changes.

To remove a skill that can be deleted, open it, choose Delete this skill, then confirm. Its instructions and associated files are removed. Closing a modified document asks you to confirm that you are discarding unsaved text.

When an MCP service declares paid tools, its card shows their estimated costs and a Confirmation threshold per conversation. A manager can change this threshold, then apply the changes. A confirmation is requested before the threshold is exceeded, even when the tools are authorised automatically. The value 0 turns off this threshold only. The amounts come from the service, not from its invoice; the counter resets when the connections restart. Services with no declared costs do not show this setting.

In Skills, three tabs separate procedures, Tools and MCP. The settings apply to the selected work profile. An authorised member can view them; only the owner and administrators can change them.

In Tools, search for a group, look at the tools it contains, turn the groups you want on or off, then choose Save tools. Connection required means that a credential or service the group needs has not been detected yet: turning the group on is not enough to configure it.

The agent can also suggest a connection to you in the conversation when it is missing a service.

In MCP, Connect a service opens a catalogue of ready-to-link services: Stripe, Notion, Linear, Airtable, Calendly, Canva, Sentry and many more. Choose a service, read what the connection allows, then Continue to that service: its authorisation page opens in the same window. Sign in to the account you want to link, accept the access requested and you come back to Diadems. Then choose Apply the changes so that the agent loads the new tools. Services marked No account are public and are added without authorisation.

The agent then uses this account in all its conversations, tasks and channels, with the rights you have granted. Approvals before action do not apply to these tools: grant only the access that is needed. Tokens stay in the agent; Diadems does not keep them and never displays them. A new authorisation replaces the previous connection, and removing the connection also deletes its authorisation. Finish the connection in the browser that started it, on diadems.fr; from the Windows and Android apps, open diadems.fr in your browser to connect a service.

For another server, choose Add an MCP and enter a name, a public HTTPS URL and, if needed, a Bearer token. The token stays in the capsule and is never shown again. When editing, leave this field empty to keep the existing token. The list of allowed tools takes one name or pattern per line; when empty, it allows all tools, subject to any advanced exclusions already defined in Hermes.

After saving, open the connection and use Test the connection to view its catalogue, without running any tool. This test validates access to the server at that moment; it does not guarantee every future action. You can edit or remove a connection. An authorisation-based connection is reconnected from its card. Local connections can still be viewed, turned on and off, and deleted; their advanced configuration is done in Hermes. Diadems services stay protected: on their cards, View tools shows, read-only, what Diadems adds to your agent, grouped by use (teamwork, monitoring, email, vault, webhooks…), with the technical name of each tool. Only the tools actually available to this agent are listed.

Saved changes are flagged as waiting to be applied. Choose Apply the changes, then confirm the restart of the connections. It affects all the agent’s profiles and can interrupt an exchange in progress. Conversations, files and memories are kept. A stopped agent is not started by these screens. If another change has been made since you opened the form, refresh before trying again.

The mechanisms used are Hermes’s own: tools and MCP servers.

Open Skills → Learn a skill, then choose a written procedure or a URL. State the steps, the expected result and any precautions. Continue in the chat prepares a new conversation; review and send the request so that the agent creates the skill with its native tools.

The Ask [agent name] button also opens a new conversation with the request “I’d like to know more about Skills” and the link to this guide, then sends it once. The agent can explain, create or edit the skills available to it. It does not edit the skills built into its image.

The native LLM Wiki skill lets you structure a body of documents. Give it an explicit task: create a wiki, ingest, search or audit sources. The agent keeps interlinked Markdown files in the wiki folder of its file space.

From Files, Add to the knowledge base prepares a task in the chat: you approve it before it is sent. The wiki stays readable and exportable. It complements conversational memory and reusable procedures.

Open Memory → Knowledge and memories. The available views are:

  • Knowledge: read the pages that memory builds automatically from memories, when there are any;
  • Memories: search facts, experiences and observations;
  • Relations: see the links between the available memories;
  • Advanced summaries: read the summaries already saved.

You can filter memories by type, browse the pages and open an item to read its detail. These views are read-only: they do not trigger an on-demand summary, a purge or an import of old conversations. They do not wake a stopped agent. An outage is flagged; use Try again once the agent is reachable.

Hermes natively handles retention and recall with Hindsight. Memories stay in the capsule; the portal reads them on demand without storing them on Diadems’ central servers. Only the digest described below, made of short excerpts, is kept there, encrypted. Memory is common to the whole agent: the privacy of a session in the portal does not isolate the facts the agent retains from it.

With each message, only the few memories closest to your request are passed to the agent, and none for a greeting, a general question or a simple “do it”. When it lacks information about a person, a file or a past decision, the agent searches its memory itself; this search appears in the steps of its answer.

So that a badly worded question does not miss what the agent knows, the agent also receives in its conversations a short digest of its memory, rebuilt every night: people, clients and partners, projects, ways of working, where to find each piece of information. Each line tells it where to look; it always checks before answering. An agent with no memories has no digest.

Tell the agent in the chat: “forget that…”. It searches for the memories concerned, removes those that match your request and lists what it removed. They no longer come up in its answers. Because memory is common to the whole agent, this removal applies to the whole team; the Diadems team can undo it for 30 days if you ask. The Memory reference notes and your conversations are not changed.

A manager sets the memory model in Settings → Intelligence, on the Memory line. Memory → Memory model opens Intelligence directly. The memory model is independent of the conversation model; its choice is kept for each connection: OpenAI subscription, Diadems credits or provider with your own key.

Edit lists the Diadems models, your OpenAI subscription and your keys; for a personal provider, type its exact identifier in the search. The reasoning effort choice appears on the line only when it is supported. Confirm and apply saves the conversation and memory settings in one go. The apply bar announces what will be reloaded: memory and the chat connection if memory changes, only the chat connection if only the conversation changes.

This reload can interrupt a reply or a task in progress. Files, memories and sessions are kept. A stopped agent stays stopped; the setting will be applied at its next start.

Fallback model is a Diadems model used only if the memory model fails: OpenAI subscription at its limit, provider error, rate limit. The memory is then saved again immediately with that model, charged to the free credit and then to the reserve. This fallback does not exist when memory uses a personal key, called directly, and it does not take over from a model that is merely too slow. On Diadems memory, changing only the fallback model does not restart the agent.

Memory → Memory status distinguishes access to memories, the connection to the provider and the recent confirmation of a native save. After a model change, a successful restart is not enough to confirm a new retention. If applying fails, check the connection and apply the setting again. The guide Agent states describes the portal’s indicators.

Adjust the organisation’s reference notes

Section titled “Adjust the organisation’s reference notes”

In Memory → Organisation’s reference notes, add or open a note, edit it, then save. These explicit facts complement automatic memory. Use short, verifiable sentences:

The weekly report must be delivered on Friday.

The counter shows the space available. Concurrent updates are detected: an outdated version does not overwrite a change made by the agent. Deleting removes the explicit note; it does not automatically erase traces of the same fact in other memories or in sessions.

Credentials and passwords belong in the Vault, never in memory or in the text of a skill.

If Diadems has turned on the TypeSafe option for your agent, it can ask for an opinion on a specific piece of information: keep, ignore or review. This tool does not save or delete any memory; nor does it change automatic memory. Its opinion can be wrong, even when it reports high confidence.

This option sends the submitted information and its context to TypeSafe, in the United States. It is off by default. Once Diadems has opened the option, the account owner or an administrator accepts or withdraws this non-EU processing in Settings → TypeSafe evaluations in the United States. Withdrawing consent stops evaluations immediately; the agent’s tools change at its next restart, or straight away with Apply now, which restarts it briefly after confirmation. Its calls use the organisation’s Diadems credits, with no rate limit: the agent can use it to sort or classify large volumes. With this consent, Diadems can also use it in its own features for this agent. Information containing a secret in a recognisable format (API key, token, password written in an address or a command) is never sent: the opinion is then “ignore”, at no cost. Still, do not give it passwords or other secrets.

With the same TypeSafe option, the agent can have a whole list or file in your file space evaluated (CSV, JSONL or JSON), for example to classify thousands of products, messages or requests against your criteria. Each row is evaluated separately. The result is a new file placed next to the original (name.jev.csv or name.jev.jsonl): it contains all the rows with, for each question, the answer and its confidence, and a jev_needs_review column that flags the rows to check (confidence below the chosen threshold, uncertain answer or row not sent). Your original file is never modified and an existing result file is not overwritten.

A run covers at most 100,000 rows and 25 MB: beyond that, the agent splits the file. It stops if consent is withdrawn, if the balance or subscription can no longer pay for it, or if Diadems restarts: rows already evaluated stay in the result file and the others are marked not_processed, so they can be run again. Only the columns useful for the decision are sent, and a row containing a recognisable secret is skipped without being sent. The answers remain opinions: check the flagged rows before acting.