RepoDaily · 2026-06-22 · ai-agent

Hermes Agent: The Self-Improving AI Agent That Learns From Every Conversation

#8 Self-hosted app Python +711 NousResearch/hermes-agent Open repository

Nous Research's open-source agent creates its own skills, remembers across sessions, runs on any model, and lives wherever you do — Telegram, Discord, CLI, or a $5 VPS.

Repo typeSelf-hosted app
Best forDevelopers and power users who want a persistent AI agent that learns their habits, creates reusable skills, and runs on their own infrastructure with any LLM provider.
Risk levelmedium
Time to evaluate30 minutes

Primary question: Do you need an AI agent that persists memory across sessions, auto-creates skills from your workflows, and isn't locked to a single model or platform?

88/100

RepoDaily adoption score

RepoDaily rates this as 88/100 (strong) for adoption: evidence, installation path, production risk, differentiation, license clarity, and AI/agent fit are scored from the article sources and adoption notes.

Directional score from RepoDaily sources and adoption notes, not a benchmark.Risk: Medium
88Evidence quality

5 source(s) across 3 source category/categories, plus a RepoDaily-specific evidence module when available.

100Installability

6 workflow step(s), 6 next-action step(s), and 3 command/install signal(s) were detected.

66Maintenance confidence

Trending momentum is +711 stars, with maintenance/release/issue signals counted when present.

93Production readiness

Risk is marked medium, with 5 security note(s) and 4 explicit skip condition(s).

91Differentiation

3 opportunity lens item(s), 4 alternative(s), and 0 type-specific section(s) support differentiation.

82License clarity

License source or license wording is present.

90Agent / AI fit

8 AI/agent-related signal(s) were detected in the article text and metadata.

Project overview

Hermes Agent is an open-source, self-hosted AI agent built by Nous Research. Its defining feature is a closed learning loop: the agent curates its own memory, creates new skills after completing complex tasks, and improves those skills during continued use. Over time, it builds a deepening model of who you are across sessions using Honcho dialectic user modeling.

Unlike most chat agents that reset between conversations, Hermes searches its own past sessions via FTS5 full-text search with LLM summarization. It runs across six terminal backends — local, Docker, SSH, Singularity, Modal, and Daytona — and communicates through Telegram, Discord, Slack, WhatsApp, Signal, and CLI from a single gateway process.

Model flexibility is a core design pillar. You can use Nous Portal, OpenRouter (200+ models), NovitaAI, NVIDIA NIM, Xiaomi MiMo, z.ai/GLM, Kimi/Moonshot, MiniMax, Hugging Face, OpenAI, Anthropic, or your own endpoint — switching is a single command with no code changes.

Problem it solves

  • Most AI chatbots forget everything between sessions, forcing you to re-explain context every time
  • Mainstream AI assistants lock you to one provider's model and ecosystem
  • Cloud-only agents don't persist on your own infrastructure or integrate with your existing tools
  • Agent frameworks require significant glue code to add memory, scheduling, or multi-platform messaging
  • No standard mechanism for agents to learn from past work and create reusable automation

How it works

  1. {'title': 'Install via one-liner', 'body': 'Run the curl installer on Linux/macOS/WSL2 or the PowerShell installer on native Windows. The installer sets up uv, Python 3.11, Node.js, ripgrep, ffmpeg, and a portable Git Bash — no admin required on Windows.'}
  2. {'title': 'Choose your model', 'body': 'Run `hermes model` to select any supported LLM provider. Alternatively, use Nous Portal for a single subscription covering 300+ models plus web search, image generation, TTS, and a cloud browser.'}
  3. {'title': 'Start the conversation', 'body': 'Run `hermes` for the interactive TUI with multiline editing, slash-command autocomplete, and streaming tool output. Or start `hermes gateway` to reach the agent from Telegram, Discord, Slack, WhatsApp, or Signal.'}
  4. {'title': 'Agent learns and creates skills', 'body': 'After complex tasks, the agent autonomously creates skills — reusable instructions plus shell commands and tool calls. These skills self-improve during subsequent use and are compatible with the agentskills.io open standard.'}
  5. {'title': 'Memory persists across sessions', 'body': 'The agent uses FTS5 full-text search with LLM summarization to recall past conversations. Honcho dialectic user modeling builds a deepening understanding of your preferences over time. Periodic nudges prompt the agent to persist knowledge.'}
  6. {'title': 'Automate with cron and subagents', 'body': 'A built-in cron scheduler delivers daily reports, nightly backups, or weekly audits to any platform in natural language. The agent can also spawn isolated subagents for parallel workstreams.'}

Six Terminal Backends, Zero Infrastructure Lock-in

Hermes ships with six terminal backends: local, Docker, SSH, Singularity, Modal, and Daytona. The serverless backends (Modal, Daytona) hibernate the agent's environment when idle and wake on demand, so costs approach zero between sessions.

This means you can run the agent on a $5 VPS for always-on presence, on a GPU cluster for heavy workloads, or serverless for bursty tasks — and switch between them without changing your agent's configuration.

Skills vs. Tools: The Contribution Architecture

  • Skills are instructions + shell commands + existing tools — the default path for new capabilities like arXiv search, git workflows, or PDF processing
  • Tools require custom Python integration, API key management, or binary/streaming handling — reserved for cases like browser automation or TTS
  • Bundled skills ship with every install; optional skills are discoverable via `hermes skills browse`; community skills live in a Skills Hub registry
  • Memory providers are now standalone plugins — the repo closed new in-tree additions to reduce coupling

Exact-Pinned Dependencies After a PyPI Worm Incident

Hermes exact-pins every direct dependency to a specific version (no ranges) in pyproject.toml. According to the project, this was tightened on 2026-05-12 in response to the 'Mini Shai-Hulud' worm hitting a package on PyPI. With exact pins, a compromised transitive package cannot reach users without an intentional version bump.

Provider-specific packages like anthropic, firecrawl-py, and fal-client live in optional extras and are lazy-installed only when the user picks that backend — reducing the blast radius of any future supply-chain incident.

Who should pay attention?

Good fit if

  • You want an AI agent that remembers your context and preferences across sessions
  • You need to deploy on your own infrastructure — VPS, GPU cluster, or serverless
  • You switch between multiple LLM providers and want one-command model changes
  • You communicate across Telegram, Discord, Slack, or WhatsApp and want one agent everywhere
  • You want scheduled, unattended automations in natural language
  • You do AI research and need batch trajectory generation for training tool-calling models

Skip for now if

  • You only need a simple chatbot API with no persistent memory
  • You require a managed SaaS with no self-hosting or infrastructure management
  • You need enterprise SSO, RBAC, or compliance certifications out of the box
  • Your team primarily uses non-Python ecosystems and prefers native language agents

Risks and cautions

Medium

Hermes is a powerful agent that can execute shell commands, manage files, and call external APIs — appropriate for technical users who understand the implications of running an autonomous agent on their infrastructure.

  • The agent runs shell commands and spawns subagents, which carries inherent execution risk on any host it runs on
  • Memory providers and skills are extensible, meaning the security posture depends on which plugins are enabled
  • Python 3.11+ is required with an upper bound below 3.14 due to Rust-backed transitive dependencies
  • Windows users may encounter antivirus false positives on the bundled uv binary
  • The project is at version 0.17.0, indicating pre-1.0 maturity with potential API changes
  • All direct dependencies are exact-pinned (no ranges) to prevent supply-chain attacks via transitive packages
  • Provider-specific packages are lazy-installed in optional extras, minimizing the default attack surface
  • Contributing guidelines explicitly prioritize security hardening: shell injection, prompt injection, path traversal, and privilege escalation
  • uv.exe false positives from antivirus are documented with attestation verification instructions
  • Memory providers must now be standalone plugins, reducing core code coupling and review burden

Alternatives to compare

ApproachWhen to useTrade-off
Claude Code
You want a coding-focused agent tightly integrated with Anthropic's Claude modelsSubscription-based
OpenHands
You want an open-source coding agent with a web UI for software development tasksOpen source, self-hosted
MemGPT / Letta
You specifically need agent memory and long-term context managementOpen source, self-hosted
n8n + LLM nodes
You want visual workflow automation with AI capabilities rather than a conversational agentFree tier available

What this trend reveals

Custom Skills Hub for Vertical Industries

Hermes's skills system and agentskills.io compatibility create an opportunity to build and share domain-specific skill packs — legal research, medical literature review, financial analysis — that extend the agent's value in specialized workflows.

Check whether the Skills Hub registry has active community uploads and whether niche skill packs gain traction in the Nous Research Discord.

Serverless Cost Arbitrage for Agent Workloads

The Modal and Daytona backends with idle hibernation open a path to near-zero-cost persistent agents. Teams could run multiple specialized agents that wake only when triggered, collapsing the cost of 24/7 agent presence.

Deploy a test agent on Modal with cron-scheduled tasks and measure monthly cost versus a fixed VPS.

Research Pipeline for Tool-Calling Model Training

Hermes's batch trajectory generation and trajectory compression features position it as a data collection tool for training the next generation of tool-calling models — valuable for AI labs and research teams.

Review the trajectory output format and test compatibility with your training data pipeline.

Best next action

Install Hermes and Test the Learning Loop

The fastest way to evaluate Hermes is to install it locally, run a multi-step task, and observe whether the agent creates and improves skills autonomously.

  1. Run the installer: `curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash`
  2. Configure your model: `hermes model` and pick a provider
  3. Run `hermes setup` to configure tools and optionally the messaging gateway
  4. Give the agent a complex multi-step task and check if it creates a skill afterward
  5. Start a new session and verify cross-session recall via FTS5 search
  6. Optionally deploy to Modal or Daytona to test serverless persistence

RepoDaily verdict

Hermes Agent stands out in a crowded agent landscape by making the learning loop — not just tool-calling — the core feature. Its model-agnostic design, six deployment backends, and multi-platform messaging gateway give it unusual flexibility for a self-hosted project. The exact-pinned dependency strategy and documented supply-chain incident response show maturity beyond its 0.17 version number. For developers who want an AI agent that genuinely grows with their workflow rather than resetting every session, Hermes is worth the 30-minute evaluation.

Sources