TL;DR
Jack Roberts demonstrates how to build a custom Claude-based AI agent system that cherry-picks the best features from Hermes, OpenClaw, and other frameworks rather than being locked into any single one. The approach gives you a persistent, memory-enabled personal AI assistant that runs 24/7 without the limitations of pre-built frameworks.
Key points
1
Hermes has 124,000 GitHub stars and is one of the most powerful AI agent frameworks available, but its limitation is that every time a better framework drops, you must re-credential all connections and lose memory continuity across systems.
2
The core strategy is 'cherry-picking winners' — paste any new GitHub repo into your coding environment (Anthropic Claude or Cursor), ask it to identify unique features vs your current system, then selectively integrate only what you want.
3
The memory system has three tiers: core memory (soul.md with personality and key facts), conversational buffer (rolling last 50-100 messages), and semantic tier indexing everything in Pinecone.
4
Self-generating skills can be added with a single prompt — after 5+ tool calls on a complex task, the agent autonomously decides if the procedure is worth saving as a reusable skill.md file.
5
Hosting recommendation is Railway over VPS for simplicity and security — it connects directly to your coding environment and runs 24/7 without exposing ports or requiring complex setup.
Actionable insights
→
To extract features from any new AI framework: copy the GitHub repo URL, paste it into Claude or Cursor, and prompt it to compare the repo against your existing system and return a prioritized list of unique features by effort level (easy win / medium / big lift).
→
Use Firecrawl for research tasks inside your agent — it scrapes live websites at roughly 80% cheaper cost than standard AI calls, and you add it simply by storing the API key in Railway environment variables.
→
For the LLM powering your agent, GPT-4.5 via an existing $20/month ChatGPT subscription is recommended because it allows OAuth authentication that Claude does not; Deepseek is noted as extremely cheap but use caution about what data you send it.
Notable quotes
“You've probably seen on YouTube, Hermes is called OpenClaw, OpenClaw is called NL. I've not seen this much killing since I watched Game of Thrones.”
“No matter how great the guys are, it's their roadmap — it's not yours.”
“You can't fix what you don't understand.”
Worth watching?
⏭️
Worth watching the full video?
The strategic framework and key technical steps are fully captured here — only watch if you want to see the live demo of prompting Claude to integrate Hermes features or the Railway deployment walkthrough.
Saved you some time? The creator still deserves a like.
Want this for your own channels?
Add the channels you follow. Every new video summarised and in your inbox the moment it drops. From £4/month.
Try it free