Jonathon's AI Wiki

Wiring Gbrain into Hermes Agent as an MCP Server

7 min read

Source: raw/Give_Your_AI_Agent_a_Second_Brain_Gbrain_+_Hermes_Agent.md (youtube.com/watch?v=-fSjdYzrFvA), Tonbi Studio (X: @tombi.studio, newsletter: onchainaigarage.com)

A practitioner walkthrough of installing GBrain and wiring it into Hermes Agent as an MCP server — distinct from GBrain’s own architecture documentation (which covers the benchmark and design claims but not a specific-harness integration) and from the brief GBrain mention in Dmitry Shapiro’s briefing (which covers the install command but not the MCP-wiring mechanics). The creator’s throughline is a clean three-way distinction between Hermes memory, an LLM-wiki, and GBrain, plus a concrete demonstration of reading from and writing to GBrain from inside a live Hermes conversation.

Key Takeaways

  • The three-way distinction. Hermes’ own memory only recalls what actually passed through a conversation with it — a session search for a topic Hermes was never told about returns zero matches, even if the information exists elsewhere (e.g. meeting notes Hermes never saw). An LLM-wiki (like this vault) is a curated, pipeline-compiled domain reference — good for “how does Hermes memory work?”-style questions, but scoped to one domain. GBrain is a knowledge base for your world — typed pages for people, companies, meetings, and decisions at any scale, and critically, it ingests things that never touched the agent’s chat (a notes folder, an Obsidian vault, meeting write-ups). The creator’s own summary line: “memory remembers our conversations, the LLM-wiki knows domains, GBrain knows our world.”
  • Architecture is a local MCP subprocess, not a service. Hermes (or any MCP-capable harness) spawns gbrain serve as a local stdio subprocess — no HTTP server, no tunnel, no token. GBrain reads/writes a local PGlite database synced from a plain markdown Git repo. Every layer is inspectable: the markdown is yours, the database is a local file, and the MCP server is a subprocess that dies with the session.
  • Install mechanics. Paste GBrain’s own “install for agent” markdown file to Hermes and let it self-install (the runtime dependency is bun). Vector search needs a separate embeddings-capable API key — a model-subscription OAuth token (e.g. Codex OAuth used for the harness itself) does not cover this. The creator uses OpenRouter since embeddings are cheap. gbrain init sets up the PGlite brain; gbrain doctor verifies the install. Optional community skill packs exist, but the install instructions explicitly say not to install them by default.
  • The MCP server is the actual integration value, not just convenience. gbrain serve wraps roughly 47 engine operations into 30+ typed MCP tools (the creator notes this had grown to 102 tools by the time of recording). Without it, Hermes has to shell out to the GBrain CLI for every call — paying full process-startup cost and reopening/closing the PGlite database each time, with every command also being another shell-approval prompt. The MCP server instead holds the database open as one long-lived process with typed tool schemas. The video’s own side-by-side (the same question asked via CLI shell-out, then via the MCP tool) shows the MCP path is noticeably faster.
  • Demonstrated bidirectional loop. Read: seed GBrain from a folder of notes, then ask “what do I need to prepare for the meeting with Alice?” — GBrain answers with cited sources (the person page, the company page, the meeting note) rather than an assumption. Write: tell Hermes “write this decision to GBrain for the Acme pilot: we start with local PGlite, they want zero infra proof of concept before considering hosted Postgres” and Hermes calls a GBrain MCP tool to create a durable, searchable decision page.
  • Graph query runs with zero LLM calls. Every page write extracts entity references and creates typed edges (e.g. worked-at, attended, mentions) via pattern matching on the links and names already present in the prose — no inference step. The creator cites GBrain’s own benchmark: the graph layer adds 31 points of precision over vector-only RAG (consistent with the +31.4pp figure documented in GBrain’s own architecture article).
  • Five rules for using GBrain with Hermes day to day (the creator’s own list): (1) tell Hermes to search GBrain first for people, companies, projects, meetings, and decisions; (2) answer from the pages themselves, not assumption; (3) write durable decisions back to GBrain as brain pages; (4) everything comes back cited; (5) memory is not the brain — world facts belong in GBrain, not in Hermes’ own memory files.
  • The two systems are additive, not competing. Hermes memory still does its job as the lean, always-loaded curator layer inside every prompt — “what did we talk about last month” still resolves via Hermes’ own session search. GBrain is the world catalog for everything that never touched a chat.
  • Best-fit audience, per the creator. Solo entrepreneurs or small teams juggling many people, companies, and projects at once, where a shared GBrain beats relying on agent memory alone.

Setup Steps (from the video)

  1. Install bun (GBrain’s runtime dependency).
  2. Paste GBrain’s “install for agent” markdown instructions to Hermes and let it self-install, or run the manual commands directly.
  3. Obtain an embeddings-capable API key (OpenRouter or any of GBrain’s supported providers) — a model-subscription OAuth token does not cover this.
  4. gbrain init (sets up the PGlite-backed brain) → gbrain doctor (verifies the install).
  5. Seed the brain: point Hermes at an existing folder of notes and ask it to seed GBrain with that data.
  6. Add the GBrain MCP server to the agent config — either ask the agent directly (“can you add the GBrain MCP server?”) or edit config.yaml manually — then reload the MCP server.
  7. Test a read (ask a question that needs cited sources) and a write (ask Hermes to record a decision to GBrain).

Try It

  • If you already run Hermes (or any MCP-capable harness) and separately keep an Obsidian vault or notes folder that never gets discussed in chat, install GBrain and seed it from that folder — the read-with-citations test (“what do I need to prepare for X”) is the fastest way to feel the difference from a plain session search.
  • Add the GBrain MCP server rather than relying on CLI shell-outs — the video’s own side-by-side showed the MCP path is noticeably faster and avoids repeated shell-approval prompts.
  • Adopt the “write decisions back” habit deliberately: after any Hermes conversation that reaches a durable decision, explicitly tell the agent to record it to GBrain rather than trusting it survives in session memory alone.
  • Budget for an embeddings-only API key (OpenRouter or similar) separately from your model subscription — the cost is small, per the creator, but it is a separate line item.

Open Questions

  • The video doesn’t specify the ongoing dollar cost of the embeddings provider beyond “not going to cost you a lot of money.”
  • Whether the MCP-vs-CLI speed difference was benchmarked with real numbers or just observed qualitatively in the video isn’t stated — treat the “much faster” claim as the creator’s informal observation, not a measured benchmark.
  • The video doesn’t cover GBrain’s “company brain” / multi-tenant mode (documented in GBrain’s architecture article) in a multi-user Hermes context — worth checking whether Hermes’ own profile isolation (Profiles & Multi-Instance) composes cleanly with a single shared GBrain.

Graph View

Backlinks