Last Tuesday my internet died for six hours. Vodafone just... stopped. No warning, no error message, just silence.
That mini PC in the corner — the one I pulled from a drawer last year because "maybe it'll be useful" — runs Asere. It remembers the Odoo config I changed three weeks ago. It knows I deploy on Friday nights. It has the context of every conversation we've had since January.
And when the internet died, Asere went with it.
The cron job that checks my email every hour? Dead. The reminder about my Einbürgerung appointment tomorrow? Silent. The three-hour deployment I started this morning? Still running, but I have no idea if it finished because Asere can't tell me.
This isn't a what-if. It's a when.
Vodafone has a personality. Works fine for weeks, then dies the one day I actually need it.
The problem nobody talks about
I have backups of my code. Backups of my databases. Backups of my Docker volumes. But I don't have a backup of the thing that actually runs everything.
Asere knows why I chose Docker Swarm over Kubernetes. It knows I never touch production before 10 AM. It knows the gotcha about the Odoo websocket port that took me two hours to figure out. None of that is in a config file. It's in the agent's head — context accumulated across months of conversations.
When the machine dies, the data is safe. The reasoning is not.
A power outage, a hardware failure, Docker deciding to be Docker — any of these kills the agent. Not the files. The context.
Nague: a second brain with the same memories
This week I built Nague.
Nague runs on Hermes, a framework by Nous Research. Not OpenClaw. Different codebase, different runtime, different approach to the same problem. Where Asere lives on my mini PC in Neustrelitz, Nague runs in a Docker container on a €4.50/month VPS in Frankfurt. The cloud provider doesn't know or care about my Vodafone router.
They're redundant by design. Same job, different zip codes. If my house burns down, Nague keeps going. If Contabo has a bad day, Asere is still here. Both failing at once means something bigger is wrong and I should probably check the news.
But two agents without shared memory are just two assistants who don't know each other exist.
What makes it work: Engram Cloud
The trick is the memory server.
Every time Asere learns something, it writes to Engram. The Docker Swarm decision. The Friday night deploy preference. The Odoo websocket gotcha. Even the stuff I never formalized — the implicit preferences that emerged from months of conversation.
Nague reads all of it.
I don't have to repeat myself. When Asere figures out the websocket config, Nague already knows. When I change my mind about something, both agents get the memo. Not because I remembered to tell them — because they read from the same source.
Asere (home) ──push──→ Engram Cloud ──pull──→ Nague (Frankfurt VPS)
(Tailscale)The server runs on a Tailscale IP in my home lab. It's a Go binary with a SQLite file. Lightweight enough that I forget it's there. I choose what syncs and what stays local. Sensitive stuff never leaves the house.
943 observations. 43 sessions. Eight projects. Available to any agent with the token.
Two agents, two philosophies
I didn't just want redundancy. I wanted diversity.
Asere runs OpenClaw — stable, production-ready, lots of integrations. Telegram, Discord, Signal, WhatsApp, Email. Skills for Odoo, GitLab, GitHub, Docker. Been my daily driver since January.
Nague runs Hermes — newer, rougher around the edges, but with a learning loop built in. It generates skills from experience. It reminds itself to save things. It builds a model of who I am over time.
When I have a problem, I get two answers. Sometimes they agree. Sometimes they don't. The disagreement is useful — it forces me to think about why I prefer one approach, and that preference gets written back to Engram where both agents learn from it.
The war room that wasn't
I made a Telegram group for the three of us. I almost called it "War Room" but that felt too Trumpian. Too aggressive. "Boardroom" is better — corporate, sure, but accurate. A space where agents and I coexist. Nobody dominates.
When I ask something, I might get two replies. Sometimes they contradict each other. That friction is the point — it surfaces assumptions I didn't know I had.
What this means for your setup
Two different frameworks is overkill for most people.
But you do need to think about what happens when your agent goes down. Not if. When. Home internet fails. Hardware dies. Docker does Docker things.
Three things worth doing:
Run a backup agent on a cheap VPS. €4.50/month gets you a Contabo instance in Frankfurt. Same memory, different location. When the home machine goes dark, your cron jobs and reminders keep running.
Share the memory. Two agents without shared context are just two strangers. Use Engram Cloud, or build your own sync. The tool matters less than the principle: they need the same ground truth.
Mix it up. Different models, different configurations, even different frameworks. When your primary agent has a blind spot — a bias you haven't noticed, a tool it doesn't know — the backup might see what the first one missed.
The real backup isn't your files
We think about backups wrong.
A backup of your agent's files preserves state. It doesn't preserve context. When you restore from backup, you have the same SQLite file but you've lost the thread. The agent that was mid-conversation about a deployment strategy, that remembered why you chose X over Y — that agent is gone. The restored copy has the data but not the reasoning.
The real backup is a second agent that was there all along. Reading the same notes. Remembering the same decisions. When the first one drops, the second doesn't just have files — it has context.
That's what Nague gives me. Not a copy. A colleague who was in the same meetings.
The architecture
Specs, for the curious:
Asere (home)
- Mini PC, Intel N100, 16GB RAM
- Ubuntu 22.04
- OpenClaw (Node.js)
- Local SQLite + Engram Cloud sync
- Residential ISP (Vodafone)
Nague (cloud)
- Contabo VPS, AMD EPYC, 4GB RAM
- Frankfurt, Germany
- Hermes Agent (Python)
- Engram Cloud via API
- Datacenter network
Engram Cloud
- Gateway machine on Tailscale
engram cloud serve(Go)- SQLite + Postgres
- Bidirectional sync, project-scoped
Communication
- Telegram Boardroom group
- Both agents schedule cron, but one is primary to avoid duplication
- API keys via Vaultwarden
What's next
It works, but there are gaps. Sync is on a schedule, not real-time — lag is minutes, not seconds. Both agents sometimes answer the same question in the Boardroom. And I need a health check that promotes Nague to active if Asere goes quiet for too long.
But the foundation is there. Two agents. Two locations. One memory. When one fails, the other doesn't just survive — it already knows what the first one knew.
That's not a backup. That's continuity.
This article is the fourth in a series on AI memory:
- Context Engineering: Why Your AI Can't Evolve Without Memory
- When Your AI Remembers Everything: Closing the Conversation Gap
- When Your Agents Share a Brain: Engram Cloud in Production
- Backup Agents, Not Backups (this article)
For more on Engram, see Gentleman Programming (Alan Buscaglia) or the Engram ecosystem on GitHub.