Hermes

A Hermes agent loads MCP servers from the mcp_servers section of its config.yaml. Set up an agent workspace and an access policy first: steps 1 to 4 of the Quickstart.

The config entry

mcp_servers:
  meaning-memory:
    command: mm-mcp
    env:
      MM_AGENT: support-bot
      MM_ACTOR: support-bot
      MM_BACKEND: file
      MM_WORKSPACE: /srv/support-bot-mm
      MM_CONFIG_DIR: /srv/support-bot-mm
      MM_SCOPE_AUTH_CONFIG: /srv/scope_auth.yaml
      MM_ALLOW_INTERNAL: "1"

Hermes starts mm-mcp as a subprocess. The variables are the same as for Claude Code and Cursor. Quote "1" so YAML passes it as text.

If Hermes runs in a container, mm-mcp, the workspace and the policy file must all be available inside the container, at the paths given here.

Give the agent only the tools it needs

Meaning Memory serves 53 tools. Most agents need a handful, and a shorter tool list leaves more of the model’s context for the task. Hermes can filter by name:

mcp_servers:
  meaning-memory:
    command: mm-mcp
    env:
      # … as above
    tools:
      include: [mm_remember, mm_search, mm_due_commitments, mm_close_commitment, mm_satisfy_commitment]

tools.exclude does the opposite: everything except the names listed.

Read-only access

Leave out MM_ALLOW_INTERNAL and the agent can search memory but not change it. The commitment tools, including mm_due_commitments, are turned off too. They still appear in the tool list and refuse each call, so also use tools.include (above) to list only the tools the agent should see. See Read-only access.

Tell the agent when to use memory

Hermes agents follow their system prompt. A few lines there make the difference between memory that is used and memory that sits idle:

Before answering a question about a customer, a past decision or something you were told, search memory with mm_search using one or two distinctive words. When you learn a durable fact, store it with mm_remember. When you promise to do something, store it as a commitment with a due_at, and check mm_due_commitments at the start of each task.

On the file backend, the advice to search with one or two words matters: search there matches the query as one phrase. See When memories become searchable.