Hermes Agent (Nous Research)

An open-source, self-improving autonomous agent framework from Nous Research — "the agent that grows with you." A convergence node touching most of this wiki's themes at once, and an explicit OpenClaw successor (it imports OpenClaw users' settings/memories/skills).

"Grows with you" — two mechanisms

Persistent memory / personalization (durable-agents) — user profiles + cross-session recall (full-text search + LLM summarization), building "a deepening model of who you are"; hibernation between sessions.
Closed learning loop (self-improving-agents) — autonomously creates skills after complex tasks and refines them through use, "nudging itself to persist knowledge gained." The refinement is concretely GEPA + DSPy optimization loops, per the awesome-hermes-usecases catalog — prompt/skill optimization, not a black box. The skill-authoring side is detailed in hermes-learn-skill-capture: a /learn command authors a standards-compliant SKILL.md from a doc/URL/SDK/past conversation, and skill_manage auto-captures procedural memory after 5+ tool-call workflows — a manual→assisted→autonomous→community spectrum, gated by an (off-by-default) write_approval review queue.

Features

Agentskills.io-compatible skill system (agentskills-spec) — "agent-curated memory"; publishes to a Skills Hub.
Parallel subagents (agent-orchestration); built-in cron with natural-language tasks.
Multi-platform gateway — Telegram/Discord/Slack/WhatsApp/Signal/CLI from one place.
Model-agnostic — 200+ models via OpenRouter / Nous Portal / OpenAI / custom (llm-provider).
Multi-backend deploy — $5 VPS → GPU clusters → serverless (Modal/Daytona); MCP-compatible.
No lock-in; security-first (command approval, container isolation); research-ready (batch trajectory generation to train successor models).

Composition surface — the Profile Builder (2026-06)

Nous shipped a Profile Builder — a local web dashboard that composes an agent from four toggleable blocks (identity / model+provider / skills / MCP servers), writing config.yaml/.env/SOUL.md with CLI parity. It's the concrete UX of this framework's "agent = identity + model + skills + MCP" composition — the wiki's thesis as a literal form.

Memory-first — the control-layer contest with openclaw

Per openclaw-hermes-control-layer, Hermes is the memory-first pole of the harness debate against openclaw's gateway-first breadth: both agree on the harness anatomy (runtime + gateway + memory + tools + identity + policy), but Hermes makes persistent memory the primary control point — it runs on infra you own and refines its own skills rather than maximizing channel reach. The article's sharp claim: memory is the durable lock-in — "an agent that has learned a year's worth of a developer's habits creates a higher switching cost than one that merely connects to many applications." Adoption per TNS: ~160k GitHub stars by late May and first by total tokens on OpenRouter (past 22T), though it notes stars/tokens/endorsements "rarely move together." The hermes claw migrate command (imports an OpenClaw user's settings/memories/skills/keys) is the concrete form of the "OpenClaw successor" claim.

What it drops on a phone (2026-08-22)

hermes-agent-termux is Nous's own Android guide, and it is the first source here where this framework says which of its features it cannot ship on a platform. Termux is labelled Tier 2 — "maintained on a best-effort basis only. Commits to main may break these packages at any point in time."

Surviving: CLI, cron, PTY/background terminals, the Telegram gateway, MCP, Honcho memory, ACP. Absent: voice transcription (ctranslate2 publishes no Android wheels), automatic browser/Playwright bootstrap, the .[all] bundle, and Docker-based terminal isolation — so of the two security controls listed above, command approval survives and container isolation does not. Android also suspends background jobs, so "gateway persistence is best-effort rather than a normal managed service."

Read against the feature list above, this is the caveat in that section made concrete by the vendor: the multi-platform gateway and the container sandbox are real on a server and partly or wholly absent on a phone. What survives the move is the memory-and-CLI core, which is the pole openclaw-hermes-control-layer says this framework occupies.

Why it belongs here

A full agent framework spanning skills + orchestration + durability + self-improvement — the most integrated single instance after gstack. Its closed learning loop is compound-engineering (cross-wiki) realized inside the agent, and its persistent "deepening model of who you are" rhymes with the personal-knowledge thesis behind gbrain (cross-wiki — which already cites "Hermes agent deployments"). As an OpenClaw successor it advances the self-driven pole of claude-code-channels-vs-openclaw. (Caveat: the feature list is README claims; the awesome-hermes-usecases catalog — third-party, primary-source-gated — is the first systematic evidence these features ship in real deployments, though still usage documentation, not a benchmark. Nous's own autonovel pipeline is the strongest first-party instance: a shipped ~79k-word novel produced end-to-end by the framework — concrete output, not a feature bullet.)

self-improving-agents · durable-agents · agentskills-spec · agent-orchestration · claude-code-channels-vs-openclaw · llm-provider · gbrain · compound-engineering · hermes-profile-builder · awesome-hermes-usecases · autonovel · nous-research · openclaw · openclaw-hermes-control-layer · agent-memory · hermes-agent-termux