3 слоя памяти · reboot ≠ wipe · матрица Pi/VPS/M4 · TCO 24 мес · 6-шаговый runbook
Hermes Agent «умнеет» не магией, а трёхуровневой памятью на диске: USER.md ~1 375 символов профиля, MEMORY.md cap ~2 200 символов на Skill, SQLite FTS5 для full-text retrieval — Skill пишется только после 5+ tool calls в завершённой задаче. Разбор архитектуры, почему нужен 7×24 uptime, бенчмарк Pi / VPS / Mac Mini M4 с измеренными RAM band, TCO 24 месяца и 6-шаговый runbook. Reboot не стирает persisted memory; sleeping Gateway рвёт каналы и Skill polish loop.
Многие видят Nous Research Hermes Agent как chat shell с tools. Persistence layer глубже. Tier 1: session context — tool state и reasoning trace в RAM, gone on restart. Tier 2: Skill Documents, markdown playbooks после complex tasks, dedup в data directory — community cap ~2 200 символов. Tier 3: persistent user model в USER.md, budget ~1 375 символов.
Retrieval через SQLite FTS5 по Skills и memory entries. Agent query local index перед context injection — дешевле по tokens, значит disk IO и index health так же критичны, как compute. Skill synthesis gate: минимум пять tool calls на task; хост не должен suspend mid-chain.
| Компонент памяти | Размер / механизм | Survives reboot? | Зачем 7×24 |
|---|---|---|---|
| Session context | Current turn + tool state | Нет — reconnect | Gateway live; IM webhook timeout = broken chain |
| USER.md | ~1 375 символов профиль | Да — on disk | Migration = data dir copy; sleep тормозит profile iteration |
| MEMORY.md / Skills | ~2 200 символов на entry | Да — on disk | FTS5 растёт с writes; backup non-optional |
| SQLite FTS5 | Local full-text index | Да — DB file | Disk jitter / VPS IO cap = retrieval latency |
Reboot ≠ memory wipe только для flushed layers. Channel UX, cron и in-flight 5+ tool chains всё равно ломаются. Субъективный 30-day narrative: 30 дней с Hermes; здесь — architecture и resource math.
Думать, reboot стирает всё: Skills и USER.md в data directory — без backup host swap = amnesia feeling.
Игнорировать 5+ tool gate: short chats не становятся Skills; sleep mid-task = extraction loop never completes.
FTS5 как black box: corrupt index / full disk → «писал, но не find» — мониторить data dir size и SQLite health.
Не trim USER.md: 1 375-char budget конечен; bloated profile dilutes preference weights.
Split Gateway от model host: dead Gateway + live cloud backend = IM drop — 7×24 = whole chain up.
Hermes = always-on agent: 20+ каналов (Telegram, Discord, Slack…) через webhooks в Gateway; cron будит subtasks; Honcho медленно refresh user model. Любой link down → missed callbacks, queue backlog, delayed Skill writes — субъективно new assistant каждую неделю, хотя Skill files растут на disk.
7×24 не ops theater — match time axis трёх tiers. Session layer хочет ms response; Skill layer — long tasks с 5+ tools; user model — weeks compounding. Closed laptop, intermittent NAS, VPS throttled neighbor IO — compounding curve flatten. Dedicated host = process survival, stable net, predictable disks как SLA, не gamble на power plug.
Memory compounds on disk, но feeling smarter = Gateway без пропусков смен — gap между 7×24 и «запущу когда вспомню».
Тот же curl -fsSL https://get.hermes-agent.org | bash — разница в memory bandwidth, disk IO, macOS-native path. Таблица: qualitative band для Gateway + Telegram + local Ollama Hermes-3 8B (intermittent inference). Exact numbers shift с quantization и channel count — для review meeting, не lab cert.
| Host | Idle RAM | Peak RAM | CPU / power | Hermes fit |
|---|---|---|---|---|
| Raspberry Pi 5 · 8GB | ≈1.5 GB headroom | Gateway alone ≈4 GB; local 8B not viable | Low-power ARM; SD IO bottleneck | API-only gateway; weak Skill compounding |
| Linux VPS 4C8G | ≈5 GB usable | API ≈6 GB; Docker +2 GB | Shared vCPU; IOPS cap | Remote SSH OK; no macOS — awkward Skills |
| Mac Mini M4 16GB | ≈9 GB usable | Local 8B + channels ≈14–15 GB at ceiling | Idle ≈12W; burst 25–35W | Native macOS; single channel + local at limit |
| Mac Mini M4 32GB | ≈22 GB usable | 8B + dual channel + cron ≈18–20 GB | Same silicon, less pressure | Production pick — room for Skill + FTS5 growth |
Unified Memory (UMA) на M4 режет CPU↔GPU copies при local inference; macOS — shortest installer + Ollama path. Pi saves watts, не держит 8B; VPS cheap rent, теряет time на RTT и IO — когда Skills + FTS5 в GB, важнее stable disk latency, чем экономия на month one.
Decision object: total cost run memory compounding 24 months — hardware, power, ops hours, upgrade anxiety, migration. Rental = CapEx → OpEx; для prod Skills + channels часто lower decision cost vs buy-plus-self-support.
| TCO dimension (24 мес) | Buy M4 16GB | Rent M4 32GB |
|---|---|---|
| Hardware cash flow | Upfront + tax; self-model depreciation | Fixed monthly × 24; RAM upgrade без box swap |
| Power (7×24) | ≈12–35W × 24h × 730 days | In service fee |
| Ops hours | Warranty, OS, outage on you | Hardware swap; remote KVM |
| Hermes data assets | USER.md / Skills / FTS5 bound to one machine | Backup → restore; wipe on return |
| Upgrade risk | M-series cadence tempts second buy | New spec at contract end |
| Opportunity cost | Hardware research vs Skill polish | Focus agent workflows + channels |
Pick RAM: API-only + one channel — 16GB; local Hermes-3 + multi-channel + cron — 32GB.
Order & access: lease ID, remote path; MDM для org.
Acceptance: Apple Silicon, disk ≥256GB, macOS on official Hermes path; disable sleep.
Install Hermes: official one-liner, hermes init; data dir + backup policy.
7×24 smoke: bind IM, long task 5+ tools; after 24h check Skill write + FTS5 retrieval.
Backup & off-board: export data dir; before lease end migrate USER.md / Skills + wipe per policy.
curl -fsSL https://get.hermes-agent.org | bash hermes init hermes model
Tip: Pin Hermes version в prod; после hermes model switch — watch 24h memory curve перед second IM channel.
Moat Hermes = three-tier memory compounding on disk — но curve нужен 7×24 Gateway, healthy FTS5, enough UMA finish 5+ tool chains. Pi/VPS pass installer, trim Skill curve on local inference или IO; Mac Mini M4 rental = hardware as predictable service.
VpsMesh Mac Mini M4 monthly rental: 16/32GB UMA, remote access, wipe-on-return. Цены: цены аренды Mac Mini M4, deploy/FAQ: центр помощи, order: оформить заказ.
Warning: Не migrate hosts + rebuild FTS5 + wipe Skill dir в один weekend — root-cause impossible. Сначала 24h Gateway stable, потом model routing или bulk memory import.
Нет. Skill Documents, USER.md, MEMORY.md, SQLite FTS5 on disk survive reboot; только session context breaks. Нужен 7×24 stable host + backups — sleeping laptop drops channels и long tool chains.
Gateway idle ~200–400MB; local Ollama Hermes-3 8B peaks 8–12GB. Parallel channels + cron — 16GB tight, 32GB Unified Memory safer. Tiers: страница цен.
Если Skill compound + channel uptime > owning silicon, 24-month rental = fixed OpEx — often beats buy+ops. Order: оформить заказ; setup: центр помощи.