Glyph Brain — An Agent-Readable Knowledge Vault

The memory layer behind a personal AI agent: a structured Obsidian/Zettelkasten vault that a human writes by hand and an LLM agent reads, searches, and extends — kept in sync by automated nightly ingestion.

Most AI assistants forget everything between sessions. Glyph Brain is the opposite: a durable, plain-Markdown knowledge base that doubles as long-term memory for the Hermes/Glyph agent. The agent reads it for context, writes new notes into it, and an ingestion pipeline keeps it fed from meetings and saved sources — automatically, every night.

This repository documents how that system is built: the taxonomy, the note templates, the ingestion pipeline, and worked examples. It contains no personal data — every example note here is synthetic (see A note on privacy).

Why this exists

A second brain is only as useful as it is consistent. If notes have ad-hoc structure, neither a human nor an LLM can reason across them. Glyph Brain solves that with three rules:

Everything is a typed note — every note belongs to a category with a fixed schema.
Everything links — notes reference each other with [[wikilinks]], forming a graph an agent can traverse.
Capture is automated — meetings and saved articles flow into the vault without manual filing, so the knowledge base grows on its own.

The result is a knowledge graph that's equally legible to a person in Obsidian and to an LLM reading raw Markdown.

The taxonomy

Glyph-Brain/Wiki/
├── Concepts/     evergreen ideas, explained in your own words
├── Sources/      articles, PRs, blog posts — ingested from raw captures
├── Papers/       academic papers with structured summaries
├── People/       who matters and why (private — not published)
├── Projects/     active work, outcomes, next actions
├── Meetings/     structured notes auto-ingested from meeting summaries
├── Maps/         Maps of Content (MOCs) that organize clusters of notes
├── Inbox/        raw captures awaiting triage into a permanent home
└── Templates/    the schema for every note type

Each category has a template with a consistent set of sections, so every Concept note looks like every other Concept note. The blank templates live in templates/.

How a thought becomes a note

        ┌─────────────┐     ┌──────────────┐     ┌─────────────────┐
HUMAN ──►│   Inbox     │────►│  Typed note  │────►│  Maps of Content │
         │  (capture)  │     │ (Concept/…)  │     │   (organize)     │
         └─────────────┘     └──────┬───────┘     └─────────────────┘
                                    │ [[wikilinks]]
        ┌─────────────┐     ┌──────▼───────┐
MEETINGS►│  ingestion  │────►│  Meetings/   │  ← nightly, from meeting-note source
SOURCES ►│  pipeline   │────►│  Sources/    │  ← from saved raw captures
         └─────────────┘     └──────────────┘
                                    │
                              ┌─────▼──────┐
                              │  AGENT     │  reads for context · writes new notes ·
                              │  (Hermes)  │  follows backlinks · cites sources
                              └────────────┘

Manual capture — a fragment lands in Inbox/, then gets promoted into a typed note and linked into the graph.
Automated capture — the ingestion pipeline pulls meeting summaries into Meetings/ and saved articles into Sources/ on a nightly cron, deduplicating and adding reciprocal backlinks.
Agent use — the agent treats the vault as long-term memory: it searches notes for context, writes new ones, and traverses [[links]] to assemble a picture across sessions.

Full mechanics in ARCHITECTURE.md and docs/ingestion.md.

What's in this repo

Path	What it is
`templates/`	The 9 note templates — the actual schemas, generic and reusable
`examples/`	Synthetic notes showing each type filled in, fully cross-linked
`docs/ingestion.md`	How the meeting + raw-source ingestion pipelines work
`ARCHITECTURE.md`	The capture → structure → agent-read loop in detail

Try the schema yourself

The templates are plain Markdown — drop templates/ into any Obsidian vault (or any folder) and start writing. Pair it with an agent that can read/search/write Markdown files and you have the same human + LLM shared memory described here.

A note on privacy

The live vault holds real meeting notes, contacts, and daily logs, so none of it is published. Everything in examples/ is invented — the people, meetings, projects, and sources are fictional, written only to demonstrate the schema. The People/ and Meetings/ categories exist in the real system but are represented here exclusively by synthetic samples.

Hermes / Glyph — the autonomous agent that uses this vault as its memory.

Author

Justin Finkelman — finktheartist · finktheartist.tech

License

MIT

glyph-brain