I replaced Hermes with this Claude System...
Jack Roberts presents a custom Claude-based AI agent system called 'Gravity Claw' as an alternative to the Hermes AI agent framework, demonstrating how to cherry-pick the best features from multiple AI frameworks rather than being locked into one. He shows how to integrate capabilities like self-generating skills, multi-tier memory, and voice-activated research using tools like Firecrawl and Railway for hosting.
Summary
Jack Roberts, who claims to have built and sold a tech startup with over 60,000 customers, opens by introducing Hermes as one of the most powerful AI agent frameworks available, noting its 124,000 GitHub stars and trending status. He outlines Hermes' key strengths: a background dreaming loop that reflects on sessions and prunes contradictions into memory, self-authored tools and skills after complex tasks, a four-layer memory system, and multi-platform messaging support across 15 services like Telegram and Discord.
However, Roberts argues that the core limitation of Hermes and similar pre-built frameworks is vendor lock-in: every time a better framework emerges, users must re-credential all connections, lose memories across apps, and conform to someone else's roadmap rather than their own. His solution is what he calls 'cherry-picking the winners' — using a coding environment like Anthropic's Claude or 'Anti-gravity' to analyze any new GitHub repo, identify unique features, and selectively integrate only the desired components into his existing custom system.
He demonstrates this live by having his agent analyze the Hermes repo, produce a feature breakdown categorized by implementation effort (easy wins, medium effort, big lift), and then integrate self-generating skills with a single prompt. He explains his three-tier memory architecture: a core memory layer (soul.md with personality and key facts), a conversational buffer of the last 50-100 messages, and a semantic tier indexed via Pinecone.
For hosting, Roberts recommends Railway over VPS for its simplicity and security, noting it doesn't expose ports to the internet and keeps the agent running regardless of whether a laptop is open. He then demonstrates creating a Firecrawl-powered research skill via voice dictation, tasking the agent to scrape Product Hunt for three-star reviews of speech dictation software. The skill is created and executed in a single session. He closes by recommending GPT 5.5 or DeepSeek as cost-effective model options and teasing a full course covering foundation setup, memory systems, and monetization.
Key Insights
- Roberts argues that the fundamental limitation of pre-built frameworks like Hermes is that users are bound to the developer's roadmap, forcing re-credentialing of all connections and loss of agent memory every time a superior framework emerges.
- Roberts claims that after five or more tool calls on a complex task, Hermes' self-generating skills system autonomously decides whether a procedure is reusable enough to save as a markdown skill file, skipping it if it's a one-off task.
- Roberts states that his three-tier memory system consists of a core soul.md personality layer, a rolling conversational buffer of the last 50-100 messages, and a semantic tier that indexes everything in Pinecone for long-term retrieval.
- Roberts claims that GPT 5.5 can be accessed through an existing $20/month ChatGPT subscription at no additional cost via OAuth authentication, whereas Claude does not offer the same arrangement — making ChatGPT more cost-effective for powering these agent systems.
- Roberts argues that Railway is preferable to VPS hosting for running a 24/7 AI agent because it is not exposed to the internet, eliminates security concerns about open ports, and integrates directly with coding environments like Anti-gravity or Claude Code.
Topics
Full transcript available for MurmurCast members
Sign Up to Access