Connect Hermes to Durmem

This guide shows how to connect Hermes to Durmem so its agent can save and retrieve knowledge from your stores. durmem setup --harness hermes connects the tools and installs shell hooks that bring project context into the conversation and check consequential terminal commands.

First, install the durmem command using Quickstart. Hermes itself must already be installed, so that hermes runs in your terminal.

Run in your terminal:

durmem setup --harness hermes

Setup backs up ~/.hermes/config.yaml (or $HERMES_HOME/config.yaml) before it changes anything, then:

adds a durmem entry under mcp_servers, between two durmem setup comment lines. Re-running setup rewrites only that block;
turns off Hermes's built-in memory provider (memory_enabled, user_profile_enabled and provider), so it does not compete with Durmem;
copies the Durmem skills into ~/.hermes/skills/ (or $HERMES_HOME/skills/);
adds Durmem shell hooks while preserving your existing hooks;
runs hermes mcp list and hermes mcp test durmem to confirm that the connection works.

Setup does not choose a store for Hermes. If config.yaml already has an mcp_servers.durmem entry that setup did not write, setup stops and leaves the file unchanged; remove that entry, or keep it and skip setup.

If config.yaml is a symlink, setup leaves it unchanged; use the manual configuration below.

Restart Hermes when setup finishes and review the first-use hook consent prompts below. Continue at Open the project in Quickstart to save and recall your first memory.

Hermes has a session-specific identity for Kanban work. Use the current conversation's native session ID when a Durmem tool asks for native_conversation_id; Durmem identifies it as hermes@<session id>. This identity goes stale when the Hermes session ends, so start a new session before doing more Kanban work.

Enable reminders and command checks#

Hermes asks for consent the first time each configured shell hook runs. Review and approve the Durmem command, durmem hook --harness hermes. Setup does not enable hooks_auto_accept or approve hooks for you. Until you approve them, the MCP tools can work without the reminders or command checks.

Gateway and other headless sessions, such as Telegram, have no terminal for the consent prompt; their hooks do not run until you approve them. Open Hermes in a terminal once and approve each displayed Durmem hook. hermes hooks has no approve command. If you have reviewed every pending hook for this Hermes profile, the documented per-run hermes --accept-hooks chat option accepts the unseen hooks for that invocation. It does not set a persistent auto-accept preference.

The first model call brings in the project's starting context; later prompts bring in relevant saved guidance. Before a consequential terminal command, the shared attestation check can block the call until the agent declares which standing memories it consulted. This checks that the agent looked, not whether the command has your approval.

Hermes currently consumes reminder text from its before-model hook, but not from its before-tool or after-tool hooks. Tool-event advisory text is therefore not delivered to the agent; ask it to retrieve relevant guidance explicitly when needed. The integration also has no pre-compaction capture hook. On Hermes versions without post_compaction, it cannot automatically re-prime after compaction: ask the agent to refresh its project context, or start a fresh session.

To inspect the configured hooks, run in your terminal:

hermes hooks list

The list should show the Durmem command for on_session_start, pre_llm_call, pre_tool_call and post_tool_call. Run hermes hooks doctor to check that the approved hooks are healthy. See Choose a store for a project to select the coding store and project scope used by reminders.

Configure by hand#

If setup cannot be used, add this entry to ~/.hermes/config.yaml (or $HERMES_HOME/config.yaml):

mcp_servers:
  durmem:
    command: durmem
    args:
      - mcp
    enabled: true

Add this block too, so Hermes's built-in memory provider does not compete with Durmem:

memory:
  memory_enabled: false
  user_profile_enabled: false
  provider: ''

This example does not name a store, so Hermes uses your configured default store. After connecting Hermes, use loom_list_stores() to check the section names, and load the one you want with loom_load_store.

For the reminders and command checks, also add these entries under hooks:. If an event already has commands, add Durmem to that event's list rather than replacing them:

hooks:
  on_session_start:
    - command: durmem hook --harness hermes
  pre_llm_call:
    - command: durmem hook --harness hermes
  pre_tool_call:
    - command: durmem hook --harness hermes
  post_tool_call:
    - command: durmem hook --harness hermes

Hermes also needs the Durmem skills in ~/.hermes/skills/ (or $HERMES_HOME/skills/) before it starts. Your Durmem server publishes them in its plugin repository, which you can clone without signing in. The Hermes bundle is the plugins/hermes/julius-loom folder of that clone, and its skills/ folder holds what Hermes needs. For a private server, replace https://durmem.ai with your server's address.

Run in your terminal:

git clone --depth 1 https://durmem.ai/durmem-plugin.git ~/durmem-plugin
mkdir -p "${HERMES_HOME:-$HOME/.hermes}/skills"
cp -R ~/durmem-plugin/plugins/hermes/julius-loom/skills/. "${HERMES_HOME:-$HOME/.hermes}/skills/"

Restart Hermes, then check the connection.

Run in your terminal:

hermes mcp list
hermes mcp test durmem

The list should include durmem, and the connection test should confirm that it responds.

Agents without hooks or a long-lived connection#

This section shows how to use Durmem from an agent whose Durmem lifecycle hooks are not wired, or that opens a new MCP connection for every command it runs. Durmem still works with such agents, with two adjustments.

A store you load stays loaded only for the MCP session that loaded it. An agent that connects anew for each command would lose the store between commands, so it should start durmem mcp once and send the load and the calls that depend on it over that one session. Each call should also carry the conversation's id as _meta.threadId. Durmem matches that id to the agent's id, <harness>@<id> (for example hermes@<conversation id>), in whoami and in Kanban owner checks. A client that cannot send it can set the harness's session environment variable instead (HERMES_SESSION_ID for Hermes).

For Kanban worker steps, ask Hermes to run echo $HERMES_SESSION_ID in its terminal and pass that value as native_conversation_id when it calls whoami. Hermes changes the id when it compresses the conversation, so read it again afterwards.

Without approved hooks, the automatic reminders do not happen: Durmem cannot surface standing context before an action or guard risky commands. Replace the first of these by hand. At the start of each conversation, and again after loading a store, the agent should read what this call returns, with the loaded store and a short description of the task in hand:

loom_read_probe_bundle(store="<store name>", query_text="what I am about to work on")

Update or remove the connection#

After running durmem update, run durmem setup --harness hermes again to refresh the skills, then restart Hermes. If you configured Hermes by hand, copy the skills again as above (git -C ~/durmem-plugin pull first).

durmem uninstall removes only the marked MCP and hook entries that setup wrote. An entry you added by hand, the memory settings and the skills setup copied to $HERMES_HOME/skills/ (or ~/.hermes/skills/ by default) stay in place; remove them yourself if you no longer want them.

Connect Hermes to Durmem

Sign in and set up#

Enable reminders and command checks#

Configure by hand#

Agents without hooks or a long-lived connection#

Update or remove the connection#