AIVM Brain connects to Hermes Agent as an MCP server, giving it governed recall over your full knowledge: projects, decisions, documents, history. Hermes ships a built-in MCP client, so the setup is one config entry. The brain works alongside Hermes's own memory files; it is independent infrastructure you own, not a Hermes component.
Add the brain to Hermes
Hermes discovers MCP tools at startup from its config. Add the brain as a server entry (Hermes supports stdio servers with command, args, and env; use hermes mcp add or edit config.yaml, per the Hermes docs).
Generate your key. Sign up free at brain.aivm.io, open Connect, pick API / MCP, and mint your agent key.
Register the server with Hermes. Add an entry under mcp_servers in Hermes's config.yaml (or use the interactive hermes mcp add). The three values Hermes needs, shown in standard MCP JSON shape:
{ "mcpServers": { "aivm-brain": { "command": "npx", "args": ["-y", "@aivm/brain", "serve"], "env": { "AIVM_BRAIN_URL": "https://brain.aivm.io", "AIVM_AGENT_KEY": "ak_your_key" } } } }Verify from a real session. Have Hermes store a fact through the brain tools, restart it, and ask for the fact back. If a fresh process can answer, the wiring is right.
Field names and the config path belong to Hermes and can change between versions; the Hermes MCP docs are the source of truth for where these values go.
Bounded files are a feature. They are also a ceiling.
Hermes Agent's built-in memory is deliberately small: as of mid-2026 it curates MEMORY.md and USER.md, bounded files injected into the system prompt at session start. That design keeps every session cheap and predictable, and for who-you-are context it works. But a few thousand characters cannot hold a year of decisions, a team's documents, or the reasoning behind last quarter's architecture change.
The brain is the deep store on the other side of that ceiling. Hermes keeps its quick notes; when a task needs the full picture, it queries the brain over MCP and pulls exactly the knowledge that clears its permissions. Small working memory, large governed recall. Same pattern your own head uses.
An always-on agent needs an accountable memory
Hermes is built to run continuously, respond on your channels, schedule its own work, and improve itself with skills. An agent with that much autonomy reading and writing your knowledge is exactly where governance stops being paperwork. With the brain, every recall and every capture Hermes makes is permission-checked and written to a tamper-evident, content-blind ledger. If Hermes did something with your data at 3am, you can see precisely what it touched.