Now Hermes Instantly Knows You

Hermes, the open agent from Nous Research, is built to remember you. It keeps track of how you like to work and picks up where you left off, so it gets more useful the more you use it.

Normally that understanding builds up slowly, one conversation at a time. But you don't have to start from zero, because you never really did. Everything that makes you you is already scattered across the apps and sites you use every day. Onairos gathers that footprint into one profile: your online persona, captured as the interests, traits, and patterns that actually define you.

We built onairos-hermes-mcp, a small add-on that hands that profile straight to Hermes. Install it, say one sentence, and about a minute later Hermes is personalized to you from the very first message: your style, your interests, and what you're working toward, with no cold start. Here is the whole flow, start to finish, in under a minute:

The Missing Piece in Hermes' Memory

Hermes ships with one of the more thoughtful memory systems in the agent world. A few ideas hold it together:

• Two files. USER.md holds a curated profile of your preferences, communication style, and workflow habits. MEMORY.md holds the agent's own notes on your environment and the lessons it has learned. Both live in ~/.hermes/memories/.
• Loaded at session start. Both files are injected into the system prompt when a session begins, so the agent opens every conversation already holding them.
• Self-managed and bounded. The agent edits its own memory through a memory tool, and deliberate character limits keep it sharp instead of letting it sprawl.
• More than the files. Every past session is stored locally with full-text search, background self-improvement reviews save new preferences and corrections as you work, and you can plug in external memory providers.

But all of that machinery shares one drawback: it starts from zero. Memory systems record what the agent observes, which means a fresh Hermes install knows nothing until you've spent days or weeks generating observations for it to record. USER.md begins as an empty file, and the only built-in way to fill it is slowly, session by session.

You, on the other hand, are not a blank slate. Onairos has already modeled you into a persona: an archetype, ranked personality traits, interests, and growth areas, built from the accounts you connect and sitting behind an API you control. The only missing piece was a bridge that hands it to Hermes on day one.

Where Onairos inserts

Two writers · one store

Hermes, natively

Session turns and the background review call the memory tool, adding entries as it learns you.

fills organically over many sessions

Onairos

Your persona written straight in through onairos-hermes-mcp.

day one · ~1 minute

→

~/.hermes/memories/

USER.md1,375 chars

MEMORY.md2,200 chars

→

Session start

Loaded as a frozen snapshot into the system prompt, so Hermes knows you before your first message.

Onairos doesn't replace Hermes' memory. It's a second writer into the same USER.md, landing on day one what the native loop would take weeks to accumulate.

What We Built

The bridge is a small MCP (Model Context Protocol) server published on npm as onairos-hermes-mcp. It exposes exactly two tools to Hermes:

• onairos_connect — creates a secure session with the Onairos backend and opens your browser to the Onairos connect page.
• onairos_sync — waits for you to finish the browser flow, fetches your personality traits, and hands Hermes a formatted persona to save into USER.md.

The package installs itself into Hermes automatically. A postinstall script detects your ~/.hermes/ directory on macOS, Linux, WSL, or native Windows, then registers the MCP server in config.yaml. No manual configuration for most setups.

The Setup, Start to Finish

Four steps, one of which is just talking to Hermes.

npm install -g onairos-hermes-mcp

1. 01

   Install the package

   The installer finds your Hermes config and adds the Onairos MCP server automatically. Restart Hermes if it was running.

2. 02

   Ask Hermes

   Say Connect my Onairos profile. Hermes calls onairos_connect and your browser opens to the Onairos connect page.

3. 03

   Complete the sync in the browser

   Sign in with Onairos, connect your platforms, and wait for the page to show Profile Synced.

   Platforms you can connect

   
4. 04

   Return to Hermes

   Say Done. Hermes calls onairos_sync, pulls your traits, and saves your persona to USER.md.

From that point on, every Hermes session starts with your persona already loaded.

Under the Hood: A Session Relay

The MCP server runs inside Hermes on your machine, but the sign-in happens in your browser, and the two can't talk directly. A short-lived session relay on the Onairos backend connects them:

1. Create. onairos_connect asks the backend for a session ID, stored in Redis with a 10-minute expiry.
2. Sign in. Your browser opens onairos.io/connect?session=<id> and runs the normal Onairos flow: sign in, connect platforms, generate traits.
3. Hand back. On completion the page posts your token to the session and marks it done.
4. Pick up. onairos_sync reads the token once, the backend deletes the session on read, then it fetches your traits over HTTPS and hands Hermes the persona to save.

Nothing sensitive lingers: you never type a password into the agent, the session self-destructs after ten minutes, and the token can be read exactly once before it's gone.

What Hermes Actually Remembers

The sync doesn't dump raw data into the agent. It hands over a compact, human-readable persona:

• Archetype — the single label that best describes how you operate.
• Top traits and personality scores — your strongest positive traits and personality dimensions.
• A summary — a few sentences describing who you are and what you care about.
• Growth areas and suggested actions — the traits you're working on and nudges toward them.
• Connected platforms — which sources the persona was built from.

Those five lines look simple, and that is the point. Condensing a person into something an agent can actually use was the hard part, and it took far more work than exporting raw data ever would. Hermes gives USER.md a budget of roughly 500 tokens. A dump of hundreds of trait scores, embeddings, and activity logs would overflow it in seconds and bury the signal under noise. So we spent most of our research effort on the opposite problem: how to pack the maximum amount of accurate personality into the smallest possible space.

Every field in the persona earns its place. The archetype captures the whole shape of a person in one word a model can reason from instantly. The ranked traits surface only the dimensions that most distinguish you, not a flat list of everything we could measure. The summary is written in plain language on purpose, because models personalize far better from prose they can interpret than from opaque numbers or vectors. The growth areas and nudges point forward, giving the agent something to act toward rather than a static label.

Accuracy is the other half of the work. Each trait is modeled from real behavior across the platforms you connect, cross-checked against several signals instead of inferred from a single post, then distilled down to what actually holds up. The design goal we kept optimizing against was a representation that stays faithful to who you are even after it has been compressed hard enough to fit an agent's memory.

That combination is what lets Hermes personalize from the very first message. It calibrates tone and verbosity to how you like to be spoken to, weights its examples and recommendations toward what you genuinely care about, and frames suggestions around the goals you are actually working on, all without a round of getting-to-know-you questions. And because the persona lives in USER.md, it persists and keeps evolving alongside everything else Hermes learns, so the personalization compounds over time instead of resetting each session.

Manual Setup

If the installer can't find your Hermes directory, wiring it up by hand takes thirty seconds. Open or create ~/.hermes/config.yaml and add:

mcp_servers:
  onairos:
    command: "npx"
    args: ["-y", "onairos-hermes-mcp"]

Save, restart Hermes, and ask it to Connect my Onairos profile.

Why This Matters

What we've shipped for Hermes is bigger than one integration. It's the pattern we think all personal agents will converge on: your persona lives with you, not inside any single agent, and agents request it through a consented, auditable handshake. Onairos models the persona once, and every agent you trust can load it in a minute.

Hermes is the first agent to get this treatment. If you're building an agent and want your users to skip the cold start, the same session-relay pattern works anywhere MCP does, so reach out.

Ready to try it? The quick-start lives at onairos.io/hermes-mcp, and the package is on npm.