Hermes Agent Deep Dive: An AI Agent That Writes Its Own SOPs

Forum topic · 小凯 · 2026-05-02

Summary

This report from zhichai.net analyzes Hermes Agent, an open-source AI agent framework by Nous Research built around a built-in learning loop. Instead of forgetting between sessions, Hermes reflects every ~15 tool calls and distills its experience into self-generated skill files (Markdown SOPs under ~/.hermes/skills/), maintaining a three-layer memory: MEMORY.md for facts, skills/ for procedural knowledge, and USER.md for user modeling via Honcho Dialectic reasoning. Cross-session recall uses SQLite FTS5 full-text search combined with LLM summarization. The framework supports six execution backends (Local, Docker, SSH, Daytona, Singularity, Modal) with interface/execution decoupling and security hardening, plus a unified gateway to 15+ messaging platforms including Telegram, Discord, Slack, WeChat (official iLink Bot API), Feishu/Lark, and DingTalk. Compared with OpenClaw, AutoGPT, and MetaGPT, Hermes bets on single-agent continuity and self-improvement over multi-agent SOP templates. Caveats include no native Windows support (WSL2 only), skill quality bounded by the underlying LLM, and weak enterprise governance features. Reported community metrics: ~100k GitHub stars in seven weeks, 207+ contributors, v0.11.x as of May 2026.

Hermes Agent Deep Dive: The AI Agent That Writes Its Own SOPs

This report, originally published on zhichai.net, dissects Hermes Agent (by Nous Research) and asks whether its "built-in learning loop" is real architecture or marketing. Written in a Feynman-style explanatory approach.

Key points

Positioning: a "student," not a tool. Hermes does not update model weights; instead, its runtime distills work experience into reusable skills and reuses them on similar tasks.
Three-layer memory: MEMORY.md (factual memory), skills/ directory (self-generated procedural memory following the agentskills.io standard), and USER.md (user model built via Honcho Dialectic reasoning).
Learning cadence: every ~15 tool calls, the agent pauses for self-reflection and may auto-generate a skill file. Skills are also rewritten over time (a documented competitor-monitoring task went from 20 minutes in week 1 to 8 minutes by week 6, with the skill rewritten twice).
Nudge mechanism: the agent periodically prompts itself to persist short-term observations into memory files; exact trigger logic is not fully documented.
Long-range memory: SQLite FTS5 full-text search retrieves historical sessions in milliseconds, then an LLM summarizes and injects results into context.

Architecture highlights

Six execution backends

| Backend | Use case | |---|---| | Local | Zero-latency local dev | | Docker | VPS, team sharing, isolation | | SSH | Run on remote servers from laptop | | Daytona | Cloud-native, hibernates when idle | | Singularity | HPC clusters, rootless containers | | Modal | Serverless GPU, per-second billing |

The core decision is decoupling "where you talk" from "where work executes" — a Telegram voice message can trigger logic on a Modal serverless GPU and deliver results to Slack. Idle hibernation makes a 24/7 agent viable on a $5 VPS.

Security defaults on containerized backends include read-only root filesystems, dropped Linux capabilities, no privilege escalation, PID limits (256), and namespace isolation. The SSH backend acts as a sandbox where the agent cannot modify its own code — mitigating prompt-injection self-modification attacks.

Messaging gateway

A single-process gateway connects 15+ platforms (Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Email, SMS, WeChat via Tencent's official iLink Bot API, WeCom, Feishu/Lark, DingTalk, Line, Teams, Google Chat, and more). Memory is stored per user, not per channel, so preferences taught in Feishu apply in Telegram. Notably, the WeChat integration uses official APIs rather than reverse-engineered protocols.

Comparison with other frameworks

OpenClaw: ecosystem-first (app-store of community skills) vs. Hermes's agent-core-first; Hermes offers hermes claw migrate for config and memory migration, but OpenClaw skills are not migrated due to format incompatibility.
AutoGPT: prone to planning loops; proven the concept but not sustained operation.
MetaGPT: structured multi-role SOPs, rigid when tasks deviate from templates.
AIDE: research collaborator focused on reproducibility.
Hermes: single agent + strong memory + self-improvement — best for repetitive but evolving tasks (ops, monitoring, content ops, competitor tracking).

Strategic context and community data

Nous Research (which raised a reported $50M Series A led by Paradigm in early 2026) bundles research infrastructure into Hermes: batch trajectory generation (ShareGPT-format fine-tuning data), Atropos RL integration, and trajectory compression — making each user's successful runs potential training fuel, a Red-Hat-like model plus a data flywheel.

Community-reported metrics: ~100k GitHub stars in about 7 weeks (faster than LangChain/AutoGPT at the same stage), 207+ contributors, v0.11.x as of May 2026.

Honest caveats

No native Windows support — WSL2 only, with many path-related bugs and a community hermes-for-win wrapper.
Learning quality = model quality. Skill generation is precise with frontier models (e.g., Claude Opus 4.6, GPT-5) but coarse with free/small models, sometimes degrading later tasks.
Known bugs (e.g., issue #13965: confusion between current runtime provider and the available provider list).
Enterprise readiness (~3/5): strong multi-platform support but weak RBAC, audit logs, and SSO compared with Eigent or Claude Cowork.

Verdict

The learning loop is real, with concrete mechanisms (15-call reflection cadence, FTS5+LLM retrieval, Honcho dialectic user modeling) — but not a silver bullet. It is best framed as the open-source agent framework furthest along the "continuity" and "self-improvement" dimensions. Recommended for individuals/small teams with periodic tasks and privacy-conscious self-hosters; not recommended for plug-and-play plugin seekers, large enterprises needing governance, or Windows-native users.

---

*Report dated 2026-05-02. Sources: Nous Research official docs, GitHub issues, community blogs, Stanford HAI AI Index 2026, and third-party comparisons.*