Hermes Agent from Nous Research is a resident assistant that compounds Skills and a persistent user model under ~/.hermes/—not a chat tab you close at night. Gateway cron, channel webhooks, and disk writes must continue while you sleep. This article covers why reboot myths, laptop sleep, and VPS latency break that story; Hermes three-tier memory; why Mac mini M4 fits 7×24 duty; a host matrix (memory and ops, not TCO); and a seven-step runbook with the official installer. Pricing on the NOVAKVM pricing page, orders on the order page, access on the help center. Reopen Hermes official links after each upstream release.
[ SECTION_01 ] // PAIN_MAP Misconceptions about restart, memory, laptop sleep, and VPS latency
- “Restarting clears the model.” Weights live in the cloud; you lose local state—Gateway sessions, SQLite WAL, in-flight tools, partial Skills. The agent feels worse because host continuity broke, not because the API tier changed.
- “Memory is only RAM.” Core markdown, Skills, and episodic indexes persist on disk under
~/.hermes/; retrieval quality needs directory durability and backups. - MacBook as production Gateway: Lid-close sleep pauses processes; mixing personal data and agent memory enlarges wipe risk on travel or resale.
- Any cheap VPS is equivalent: Dockerized Hermes works, but RTT to messaging APIs and Git adds jitter. Public posts often cite 2 vCPU / 4GB RAM as a Gateway floor—not an official SLA; re-verify after release.
- Deferred rsync: Without scheduled exports, episodic SQLite and Skills diverge; migration must be a planned cutover.
- Uptime vs GUI login: You need supervised Gateway service and durable disk, not a logged-in desktop 24/7.
Official docs cover VPS and SSH backends; many teams still prefer the documented macOS curl install. Links below are authoritative—reopen after each release.
Hermes Agent official documentation (Nous Research)
NousResearch/hermes-agent (GitHub repository)
hermes-agent.org (official site)
[ SECTION_02 ] // MEMORY_TIERS Three-tier memory: session, Skill Documents, and persistent user model
Hermes implements a closed learning loop: finished work becomes reusable assets retrieved in later sessions. All three tiers live under ~/.hermes/ on a host you control.
Session memory is the live Gateway surface—tool calls, channel bindings, short-horizon context. It dies with the process; laptop sleep silences Telegram or Discord even when the API still answers.
Skill Documents are procedural files—checklists, scripts, API shapes—that you can diff in Git. They compound value per successful run; ephemeral containers without volumes reset that curve.
Persistent user model (Core) stores identity and facts—often USER.md, MEMORY.md, and episodic SQLite with FTS5 per official Memory docs. Losing Core feels like personality wipe while model weights stay the same.
Snapshot ~/.hermes/ before OS upgrades, hardware swaps, or lease checkout. Reinstalling the binary without the directory is the usual self-inflicted amnesia.
[ SECTION_03 ] // WHY_M4 Why Mac mini M4: unified memory, native macOS install, low power at 7×24
Mac mini M4 fits memory-centric Hermes when you want macOS-native install, optional on-device models, and Apple automation on one box.
Unified memory (UMA) pools CPU, GPU, and Neural Engine—shorter copy paths for local tools than typical disaggregated VPS layouts. Measure your own latency and RSS; do not copy unlabeled benchmark posts—re-verify upstream docs after release.
Native install is usually one curl pipeline per docs. Low idle power makes 7×24 realistic while a MacBook travels. Use 16GB for API-heavy routes; 24GB/32GB for local models or large episodic stores—confirm SKUs on Apple and rental pages.
Monthly bare-metal rental preserves the same ~/.hermes/ semantics without CapEx on a short pilot. For dollars, see buy vs rent TCO.
[ SECTION_04 ] // HOST_MATRIX Host comparison: Raspberry Pi vs VPS vs NOVAKVM bare-metal Mac mini
The matrix below compares memory durability and operational fit for Hermes Gateway hosts. It is not a price table; verify dollars on vendor sites and the pricing page separately.
| Dimension | Raspberry Pi class (ARM SBC) | Generic Linux VPS | NOVAKVM bare-metal Mac mini M4 rental |
|---|---|---|---|
| ~/.hermes/ durability | SD or USB wear; sudden power loss risks filesystem corruption without careful journaling and backups | Block storage survives reboot; risk shifts to snapshot discipline and provider migration | Local SSD on dedicated hardware; you own backup cadence, platform documents checkout wipe |
| 7×24 Gateway realism | Possible with tuning; thermal and I/O bottlenecks under SQLite growth | Strong uptime if sized; noisy neighbors and hypervisor maintenance still pause VMs | Physical Mac tuned for always-on; no neighbor VM stealing memory bandwidth |
| Skill + episodic growth | Limited RAM and slow storage cap FTS5 and large Skill trees | Scale disk vertically; cross-AZ latency still hits tool loops | 16–32GB UMA tiers; fast local disk for episodic SQLite and Skill corpora |
| macOS-native installer and Apple toolchain | Not applicable; Linux-only paths | Docker or SSH workflows per docs; not a substitute for Xcode-on-Mac tasks | Documented curl install on macOS; same host for iOS or macOS automation |
| Ops burden | You maintain board, power, imaging, and off-site backup | You maintain OS patches, firewall, secrets, and provider lock-in exit | Remote hands for hardware; you focus on Hermes config, Gateway, and memory exports |
| Typical fit | Hobby alerts, ultra-light Skills, tolerant of manual recovery | API-only Hermes, Linux-first teams, multi-region VPS already standard | Production Gateway, compounding ~/.hermes/, Apple ecosystem adjacency, documented lease exit |
If your success metric is “the agent remembers me next month,” optimize for disk integrity, supervised uptime, and clean host migration before you optimize monthly rent alone.
[ SECTION_05 ] // RUNBOOK Seven-step runbook: rented Mac mini, Hermes install, Gateway, backup
- Define continuity requirements: List channels (Telegram, Discord, Slack, others per docs), expected
~/.hermes/growth per month, and maximum acceptable Gateway downtime during macOS updates. - Select memory tier on the pricing page: Start with M4 16GB for cloud API routes; choose M4 24GB or higher when local models or large episodic stores are in play. Confirm rates on the pricing page before checkout.
- Order and provision access: Complete the order page flow; use the help center for first login, disk free space checks, and firewall expectations.
- Create a dedicated macOS user for Hermes: Separate operator account from personal Apple IDs when multiple people touch the lease.
- Install Hermes Agent: Run the official installer under that user; verify version and directory layout immediately after install.
- Configure Gateway, providers, and channels: Wire model providers, message Gateway health checks, and log rotation; schedule tarball or rsync backups of Core markdown, Skills, and episodic SQLite.
- Document cutover and checkout: Before lease end, export
~/.hermes/to your next host, restore, smoke-test retrieval, then follow platform checkout wipe so USER profiles and Skills do not remain on shared infrastructure.
curl -fsSL https://get.hermes-agent.org | bash
hermes --version
ls -la ~/.hermes/
Install command source: Hermes Agent official site and documentation. If upstream changes the installer URL, flags, or package names, treat the docs as authoritative and re-run verification after release.
[ SECTION_06 ] // HARD_FACTS Hard facts, sources, and NOVAKVM conversion
- Hermes positioning: Open-source MIT agent stack from Nous Research emphasizing cross-session memory, Skill generation, and multi-channel Gateway rather than single-shot chat. Source: official documentation and GitHub README; re-verify after each release.
- Default memory root: User-local
~/.hermes/with Core markdown, Skill libraries, and episodic SQLite as described in official Memory sections linked above; host migration requires directory-level backup. - VPS sizing folklore: Public deployment articles often mention 2 vCPU / 4GB RAM as a pragmatic Gateway minimum on Linux VPS; treat that as community guidance, not a Hermes performance guarantee, and re-read upstream docs post-release.
- Apple Silicon UMA: Shared memory pools benefit local tool and optional on-device model paths; measure your own workloads instead of copying unlabeled throughput claims from social posts.
- NOVAKVM semantics: Dedicated Mac mini M4 and M4 Pro bare-metal leases across multiple regions for 7×24 Gateway workloads; pricing and regions on the pricing page, provisioning on the order page.
Treat the host as a memory appliance: always on, backed up, migrated as a unit. To compound Skills without day-one CapEx, rent dedicated Mac mini bare metal, run the installer above, and prove Gateway uptime across two releases before you buy.
Match tier to Skill growth on the NOVAKVM pricing page, provision from the order page, and use the help center for access. NOVAKVM Mac mini bare-metal rental keeps ~/.hermes/ on real Apple Silicon with documented checkout wipes—without making your laptop a 7×24 server.