This is a submission for the Hermes Agent Challenge
My mom sends WhatsApp voice notes I never open. She sees the blue checkmarks, waits, sends a follow up: "beta?" Then I reply at 11 PM with "sorry busy day, all good" and she goes to bed wondering if I'm actually okay.
This has been our pattern for two years. I'm 25, she's 58, we live 4,000 miles apart. The distance isn't geographic anymore. It's attentional. I have the time. I just don't have the bandwidth to perform a relationship on demand.
What I Actually Built
Not a chatbot pretending to be me. I tried that in 2024 with a GPT wrapper and felt sick for a week. What I built with Hermes Agent is stranger: a translator that lives in my work, reads my actual days, and turns them into a language she understands.
Every evening at 8 PM her time, she gets a 90 second WhatsApp voice note. Not from me. From my number, labeled clearly as "Asef's assistant." She can reply directly. It comes back to me.
The stack (May 2026):
-Hermes Agent on a €5 Hetzner VPS in Frankfurt
-WhatsApp Business API via 360dialog
-Inputs: my Google Calendar, GitHub commits (with human readable messages I actually write), Strava, and a manual "mood tag" in Notion each evening
-Output: a voice note in Hindi English mix, the way she actually talks
The prompt isn't "summarize my day. It's: "You're my mother's neighbor who saw me at the market. Tell her what I was doing, ask the questions she'd ask, use her vocabulary."
I spent three evenings feeding Hermes old WhatsApp conversations between us. Not to mimic me. To learn how she asks about things. The output reads like this:
"You were in that 'architecture review' meeting for two hours. Did they finally decide on the database thing you were worried about? Also you only ran 3k, lazybones. What are you reading? Last week you mentioned that book about trees."
She wrote "lazybones." Hermes learned it.
Week One: The Performance
Day 1: She replies "Yeh kya hai?" (What is this?) then forwards it to my dad with a laughing emoji.
Day 3: She asks the bot about my knee. I mentioned "knee sore" in a Strava comment three days ago. She remembered. The bot remembered. I didn't.
Day 5: She starts talking to it. "Tell him his cousin got engaged. He won't reply to me but maybe he'll reply to his robot."
Day 7: I open my phone to four voice notes from her, all addressed to the bot. One is her crying because her sister's biopsy came back. The bot transcribes it, flags "biopsy" and "crying," and I call her within ten minutes for the first time in months.
The Incident (May 12, 2026)
I had a bad deploy at 2 AM. My commit message: "fuck this, rewriting auth from scratch." The bot includes this in her evening update, verbatim, with context: "He had a rough night fixing something important. He used a strong word. He's frustrated but safe."
She replies at midnight her time, laughing: "Your robot told me you are stressed. Are you eating?"
I should have been embarrassed. I was, briefly. But then I realized: she called. She knew something real about my day. Not the sanitized version I give when I finally reply. The actual texture. The 2 AM despair. The thing I never tell her because I don't want her to worry. Which means I tell her nothing. Which means she worries more.
I didn't fix the filter. I started writing better commit messages. Not sanitized. Just translatable.
What the Bot Actually Does
It's not keeping my mom updated. It's keeping me updated on my mom.
Every Sunday morning, Hermes sends me a pattern report I didn't ask for but now depend on:
-"Her average message length dropped 15% this week. Last time this happened: March 2026, when she had the flu."
-"She mentioned 'lonely' twice, both times after 9 PM. Previous 30 days: zero mentions."
-"She asked about your marriage prospects again. This is the sixth time in 2026. Trend: accelerating."
I don't act on all of it. But I know it now. I know things about my mother's emotional state that I didn't when we talked directly, because direct conversation has performance in it. This has data, and data doesn't perform.
The Uncomfortable Truth
Some mornings I read her transcripts and feel nothing. The translation layer works too well. I'm reading my mother's loneliness like a Grafana dashboard. Other mornings a phrase she uses hits me so specifically that I miss her viscerally, and I realize I've been missing her for years, not just since I moved.
The bot didn't fix our relationship. It made the brokenness visible, which is harder and more useful.
What I'd Do Differently
The voice: I started with text messages. She ignored them. Voice notes in her own cadence changed everything. Text is for information. Voice is for presence.
The asking module: I built a second agent that messages me once a week: "You haven't asked your mom a question in five days. Her last message mentioned a doctor's appointment. Consider: 'How did Tuesday go?'" This isn't AI doing emotional labor. It's AI prompting me to do my own. The difference matters.
The off switch: She can text "STOP" and the agent dies instantly. She hasn't used it. I don't know if that's trust or politeness, and I'm afraid to ask.
For the Challenge
If you're building with Hermes, build toward the problem that makes you slightly ashamed. Not "I automated my calendar." Everyone does that. Build the thing that reveals something you didn't want to know about yourself.
My mom's bot is still running. I still miss 60% of her direct messages. But now I listen to her voice notes every morning, and I reply more often, and sometimes I call her on Wednesdays for no reason.
The bot didn't solve my problem. It made the problem solvable, by making it impossible to ignore.
Technical Appendix
Model: Hermes 3 via OpenRouter, switched to local Qwen 3 after May 2026 rate changes
Voice synthesis: ElevenLabs with a cloned sample of my own voice (disclosed to mom, she prefers it to robotic TTS)
Memory: Custom Hermes skill tracking "mom isms" and emotional baseline since March 2026
Cost: €5 VPS + 4/month API calls + 11/month ElevenLabs
Privacy: All voice data encrypted at rest, explicit consent documented, auto delete after 90 days
Top comments (0)