Overview
Hermes Brain: a persistent memory vault for Hermes Agent and OpenClaw.
Give your Hermes / OpenClaw agent a memory it can’t forget, and you can actually read.
Hermes Brain is an Obsidian vault template. Every session your agent runs is archived as plain Markdown, deduplicated, secret-scrubbed automatically, and staged for you to review before anything becomes permanent. There is no black box and no vendor lock-in. It is just a vault, so you can search it, link it, graph it, and back it up like any other notes.
Why this instead of nothing?
- You stop losing context. Every Hermes session is archived automatically. Nothing lives only in a chat log you’ll never scroll back to.
- You stay in control of what becomes “memory”. Nothing is promoted to Hermes’s real memory without a human reading and approving it first.
- It’s just Markdown. Open it in Obsidian,
grepit, put it in git, read it in Notepad. No proprietary format and no export step. - Set up once, in one paste. A single prompt in your Hermes chat installs the vault and wires up hourly archiving.
What it solves
| Your pain point | How Hermes Brain fixes it |
|---|---|
| AI agents can’t access your knowledge | A bundled MCP server connects Hermes and Obsidian to LLMs |
| Manual memory triage is slow | promote_memory.py --review lists candidates with a full note preview |
| Graph view is unusable noise | Colour-coded PARA folders and size-by-relevance make the graph readable |
| Installers overwrite your profile | The installers skip 02-Areas/User-Profile.md if it is non-empty |
| Documentation is scattered | A structured PARA vault with ADR, project and research templates |
How it compares
| Hermes Brain | Raw chat logs | Vector-DB memory (e.g. Mem0) | Generic Obsidian PKM | |
|---|---|---|---|---|
| Human review before “memory” | Yes, staged in 04-Archives/Memory-Review/ | No structure | No, auto-embedded | No agent pipeline |
| Readable without special tooling | Yes, plain Markdown | Yes, but unstructured | No, needs the vendor’s UI/API | Yes |
| Portable, no lock-in | Yes | Yes | No | Yes |
| Built-in agent-usage analytics | Yes, token usage, vault audit, agent performance | No | Varies | No |
| Zero setup for a Hermes agent | Yes, one-paste install prompt | N/A | No, separate integration | No |
Use a vector-DB memory service if you want the agent to auto-recall semantically similar things with no human in the loop. Use Hermes Brain if you want to see and approve what your agent remembers, in a format you already own.
The core pipeline
04-Archives/Daily/YYYY/MM/DD/*.md → 04-Archives/Memory-Review/*.md → Hermes MEMORY.md / USER.md
(raw session archive) (staged candidate facts, (durable, injected every
human-reviewed) turn, promoted by hand)
Read the Architecture page for each stage in detail.
Requirements
- Obsidian (free)
- A Hermes Agent install, or OpenClaw, if you want the automated session-archive and memory pipeline
- Optional, for the trend-digest scripts: Ollama or Python 3 with
sentence-transformers
Where to go next
- Quickstart: install the vault in about a minute
- Vault structure: the PARA layout
- MCP server: the tool reference
- FAQ