Video summary

Hermes Agent Fundamentals In 29 Minutes

Main summary

Key takeaways

Technology

Hermes Agent Fundamentals (29 minutes)

The video is a guided setup and “fundamentals” walkthrough for Hermes Agent (a local-first/multi-agent assistant). The central theme is that correct setup enables a sustainable, privacy-preserving agent experience that can become more capable over time through tools, skills, cron jobs, and a multi-tier memory system.

Core goals of the guide

  • Explain how to choose hardware for Hermes (local machine vs VPS vs any computer + Docker).
  • Walk through installing and configuring Hermes (desktop app or terminal).
  • Cover Hermes major features:
    • Tools (prebuilt capabilities like web search, image generation, browser automation)
    • Skills (instruction “modules” the agent can learn/use, and even create new skills)
    • Cron jobs (scheduled/recurring automated tasks)
    • Memory management (multi-tier memory + optional enhancements)
  • Show how to use open-source/local models (Ollama / llama.cpp).
  • Demonstrate what you can build next using multi-agent workflows, including Discord-based orchestration and Kanban coordination.

Hardware choices (where Hermes runs)

The speaker runs Hermes 24/7 primarily on:

  • Mac Studio: 64GB RAM, M4 chip

They outline four options:

  1. Dedicated powerful local machine (example: Mac Studio)
  2. VPS (cloud virtual server; can be cheap and 24/7)
  3. Use an existing laptop/PC by “wiping it” and running as a quasi-dedicated box (example: 16GB RAM start)
  4. Personal computer only if needed, but recommended to use Docker to contain the agent

Optimization point: run cheaper/local models by default, and only use cloud models when necessary.

Installation & initial configuration

  • Download/setup starts from hermesagent.newsresearch.com
  • Supports installing via:
    • Desktop app (shown as preferred for non-terminal users)
    • Terminal (also straightforward)

In the app:

  • Connect model providers (e.g., Anthropic/OpenAI) or go fully local.
  • If no provider subscription exists, an “easy option” mentioned is a monthly “News Portal/news portal” model subscription with access to many models.

Model “driver” guidance

  • General consensus: Sonnet 5 is considered a strong driver, but it can be expensive due to API-key requirements.
  • Speaker’s approach: run an open-source local model as the primary driver (example given: Qwen 3.5/5.6 35B variant—subtitles indicate “Quant 5.6 35B” / “Qwen 3.6 35B” at different points).

They also promise and reference a free guide for deeper model selection, including “exact prompts and workflows” to customize the agent based on budget and use case.

Memory: how Hermes “remembers” (and how to improve it)

The guide spends substantial time on memory because prior agents reportedly became slow, inconsistent, and “forgetful.”

Default memory system (2-tier)

Stored on the machine (macOS example):

  • .hermes/ folder
  • Two main human-readable markdown files:
    1. memory.md: high-level setup/config context (where things live, integrations, in-progress info)
    2. user.md: user profile knowledge (“human readable”, stored under memory / user data)

To retrieve older context selectively, Hermes uses:

  • Session search tool to search logs from archived conversations
  • Logs stored in a SQLite database (“sequel light” in subtitles)
  • Hermes only pulls from the archive when needed to reduce bloat/slowdowns

Memory enhancement options

The speaker describes additional tiers:

Tier 3 (optional plugin): external memory provider “Honcho’s”

  • Adds an additional layer by learning from interactions to form “implicit understanding” (patterns about the user, projects, behavior).
  • Does not replace tier 1 and 2; it supplements them.
  • Setup is initiated from a Hermes prompt like: “Help me set up Honcho.”
  • Verification via terminal commands like: Hermes Honcho status

Tier 4 (second brain): Obsidian vault integration

  • Hermes can save and access information stored in the user’s Obsidian vault.
  • Hermes uses an Obsidian skill to:
    • Save research/work artifacts into the vault
    • Later retrieve documents (e.g., PRDs) and use them for new planning or implementation steps

The speaker demonstrates retrieving a saved PRD (example: “food tracker PRD”) and generating a step-by-step MVP plan from it.

Major Hermes features

1) Tools

“Tools” are pre-configured infrastructure managed by Hermes/News Research creators. Examples:

  • web search
  • image generation / text-to-image
  • browser automations

Tools appear under Tools / Toolkits, and the speaker shows an example of searching AI chip developments.

1a) Integrations (connect third-party services)

Integrations allow Hermes to call external software such as:

  • model providers (News Portal, Anthropic, OpenAI, MiniMax, etc.)
  • messaging platforms (Discord shown)

The speaker uses Discord for alerts and daily briefs (example daily news brief audio/message).

If a native integration doesn’t exist:

  • Use MCP (Model Context Protocol) to add third-party tools
    • Example given: Notebook LM MCP for research and creating podcast-like briefings

2) Skills (instruction manuals; can learn/grow)

Skills are reusable task procedures/instructions.

Prebuilt skill categories include:

  • Apple Notes, Apple Reminders, iMessage
  • Research tools (search papers)
  • Creative content generation (example: ASCII video generation)

A major emphasis: Hermes can develop alongside you.

  • The speaker demonstrates asking Hermes to evaluate business ideas with added criteria.
  • Then prompts Hermes: “Make this into a skill…”
  • Hermes generates a new skill (example: Business Idea Evaluator) describing:
    • dimension scoring (1–10)
    • weighting
    • weighted total out of 10
    • go / caution / no go recommendation

They then show invoking it with a slash command (e.g., /business idea evaluator) and seeing computed scores and recommendations.

3) Cron Jobs (scheduled automation)

Cron jobs run recurring tasks automatically. Examples:

  • daily scheduled news brief (also turned into an audio file)
  • summarizing Apple Notes and compiling into Obsidian
  • nightly health checks to ensure hosted models and multi-agent workflows are running correctly

4) Open-source/local AI models (optional, but recommended for cost/privacy)

The speaker notes:

  • Cloud API usage is feasible but can be expensive (“hundreds of dollars” sentiment).
  • Local models are recommended for privacy and cost reduction.

Local setup options:

  • Ollama (easy model package manager)
  • llama.cpp (more direct control; mentioned as faster and more customizable)

They recommend:

  • download Ollama via ollama.com
  • use Hermes to help fetch/open-source models
  • or use llama.cpp directly after around 30 minutes of terminal setup

What you can build with Hermes multi-agent systems (demo)

The last section demonstrates building software/workflow apps using Hermes + Discord.

Example: building a Pomodoro companion desktop app

The speaker uses:

  • Discord-based multi-agent orchestration
  • Hermes drafts PRD/spec
  • Hermes queues a build
  • Build success appears in a channel
  • The app is delivered as a floating always-on-top desktop widget

They test it by creating a 1-minute timer with a label, then confirm:

  • the app overlays on top
  • completion state updates
  • data is saved to the Obsidian vault

Parallel multi-agent work

Because it’s multi-agent:

  • Hermes can work on multiple spec/build tasks in parallel
  • e.g., generate additional widget ideas while also adjusting UI theme (pastel color changes)

Orchestration control (Kanban)

They mention native orchestration via a Kanban board to coordinate multiple agents for optimized workflow management.

Main speaker / source

  • Main speaker: Tina (referred to as “Tina” throughout; shown as the user profile and demo owner of the setup)
  • Source/tool ecosystem referenced: Hermes Agent / “News Research” (creators of Hermes), plus related integrations:
    • Ollama / llama.cpp
    • Honcho’s (external memory provider)
    • Obsidian (second brain)
    • MCP (Model Context Protocol)
    • Notebook LM MCP

Original video