Agentic AI architecture and knowledge automation for an intelligent second brain by Austin BC.
MIMIR is the public engineering case study and reference implementation for AustinOS, my private production knowledge platform. It is a production-grade agentic knowledge system that serves as my intelligent second brain, orchestrating Obsidian, Codex, Hermes Agents, and Git across governed pipelines for automated ingestion, synthesis, validation, versioning, and retrieval.
MIMIR is inspired by Andrej Karpathy's LLM Wiki pattern: instead of reconstructing knowledge from raw documents at every query, an LLM incrementally compiles and maintains a persistent, interlinked Markdown wiki. Rather than implementing the operating schema around Claude Code and CLAUDE.md, MIMIR uses Codex with AGENTS.md as its control contract, Obsidian as the graph-based knowledge interface, and Git as the auditable state layer.
This repository intentionally contains architecture, sanitized configuration, reference scripts, operational evidence, and implementation notes—not the private journal, CRM, conversations, credentials, or raw source corpus from AustinOS.
AustinOS treats personal knowledge as a maintained software system rather than a passive notes folder:
- New material lands in an immutable
raw/source layer. - An agent reads the source and follows the operating contract in
AGENTS.md. - Source summaries and reusable concept, entity, workflow, synthesis, or comparison pages are created in
wiki/. - New pages are connected with Obsidian links and grounded with original-source and raw-note provenance.
index.mdis refreshed as the content-oriented navigation surface.log.mdreceives an append-only record of the ingest or maintenance event.- The result is committed to the private AustinOS Git repository.
- Hermes provides an agent runtime and messaging gateway, including an allowlisted Telegram channel configuration for remote interaction.
The production AustinOS snapshot documented in August 2026 contained:
| Signal | Observed scale |
|---|---|
| Compiled wiki pages | 129 |
| Preserved raw/source files | 1,374 |
| ChatGPT conversations represented in the corpus | 473 |
| Full conversation Markdown renderings | 470 |
| Codex ingestion schedule | Daily at 23:00 Asia/Manila |
| Git history | Nine repository snapshots from initial vault through recurring ingests |
| Agent gateway | Hermes user service enabled and active |
These are point-in-time measurements from the private AustinOS repository, not hard-coded product limits.
flowchart LR
capture[Web clips / notes / exports / transcripts] --> raw[raw/<br/>immutable sources]
raw --> agent[Codex ingest agent<br/>AGENTS.md contract]
agent --> wiki[wiki/<br/>compiled knowledge]
agent --> journal[journal/<br/>private reflection]
agent --> crm[crm/<br/>relationship memory]
wiki --> index[index.md<br/>navigation]
wiki --> log[log.md<br/>append-only chronology]
journal --> index
crm --> index
index --> obsidian[Obsidian<br/>links + graph]
log --> git[Private Git repository<br/>versioned state]
obsidian --> hermes[Hermes agent runtime]
hermes --> telegram[Telegram<br/>allowlisted remote interface]
The design separates four concerns:
- Evidence: immutable sources and canonical links.
- Compilation: agent-generated, cross-linked knowledge pages.
- Navigation: human-readable index and chronological log.
- Access: Obsidian locally, Git for state history, and Hermes for conversational access.
Read the full architecture and ingestion pipeline.
.
|-- .github/workflows/validate.yml # CI for the public reference implementation
|-- docs/
| |-- architecture.md # Components, boundaries, and state transitions
| |-- automation.md # Codex schedule and Git-safe execution model
| |-- data-model.md # Raw, compiled, navigation, and private layers
| |-- hermes-telegram.md # Hermes gateway and Telegram channel design
| |-- implementation-notes.md # What was built, observed, and sanitized
| |-- ingestion-pipeline.md # Detailed lifecycle and failure handling
| |-- privacy-model.md # Public/private boundary and threat model
| `-- screenshots.md # Evidence catalog
|-- examples/
| |-- automation.example.toml # Sanitized Codex automation definition
| |-- hermes-config.example.yaml # Secret-free Telegram/gateway shape
| |-- hermes-gateway.service # Persistent user-service template
| `-- vault/ # Minimal test vault
|-- scripts/
| |-- check_markdown_links.py # Repository-local documentation link checks
| |-- snapshot_metrics.py # Point-in-time vault inventory
| |-- vault_audit.py # Provenance, links, index, and log checks
| `-- public_preflight.py # Secret/private-path release guard
|-- SECURITY.md
`-- README.md
The scripts use only the Python standard library.
python scripts/vault_audit.py examples/vault
python scripts/snapshot_metrics.py examples/vault
python scripts/check_markdown_links.py .
python scripts/public_preflight.py .vault_audit.py checks:
- required vault directories and control files;
- wiki frontmatter fields;
- provenance sections;
- unresolved Obsidian links;
- index coverage;
- append-only log heading shape.
public_preflight.py rejects common secret patterns, private vault directories, private keys, credential files, and accidental AustinOS corpus copies before publication.
The active Codex automation is configured with this schedule:
RRULE:FREQ=WEEKLY;BYHOUR=23;BYMINUTE=0;BYDAY=SU,MO,TU,WE,TH,FR,SA
Although expressed as a weekly recurrence with all seven days selected, the effective behavior is daily at 23:00 local time. Each run checks for loose raw sources, performs the ingest contract, updates navigation and chronology, and commits/pushes only the relevant result when changes exist.
See automation.md for idempotency, dirty-worktree handling, network failures, and recovery behavior.
The documented production environment includes:
- Hermes Agent v0.13.0;
- Python 3.11 runtime and isolated virtual environment;
- OpenAI Codex provider with GPT-5.5 selected at capture time;
- persistent
hermes-gateway.serviceuser service; - messaging tools and
python-telegram-botdependency; - a Telegram configuration section with
allowed_chats,channel_prompts, andreactionscontrols; - an AustinOS vault-protocol skill available to the agent.
Credentials, chat identifiers, device codes, environment files, and personal prompts are deliberately excluded. See hermes-telegram.md.
| System surface | Evidence |
|---|---|
| Obsidian knowledge graph | |
| Codex ingest execution |  |
| Hermes agent runtime |  |
| Compiled source-backed wiki |  |
Additional screenshots and their interpretation are cataloged in screenshots.md.
Public screenshot assets are privacy-safe reconstructions: they retain the operational layout and technical meaning while using fictional usernames, hostnames, paths, and note titles.
The private AustinOS repository is the source system. It contains personal notes and other sensitive source material and is not suitable for public release. This repository uses a clean-room public boundary:
- no raw private corpus;
- no journal or CRM records;
- no exported conversations;
- no environment files, tokens, chat IDs, or device codes;
- no direct copy of private source content;
- only sanitized schemas, operational patterns, scripts, metrics, and screenshots approved for the case study.
Read privacy-model.md and SECURITY.md before adapting this design.
- Author: Austin BC
- Architectural influence: Andrej Karpathy's LLM Wiki pattern
- Public project: MIMIR
- Private production system: AustinOS (
xntr99/AustinOS, intentionally private) - Primary interface: Obsidian
- Knowledge compiler and maintenance agent: Codex
- Versioned state: Git/GitHub
- Conversational gateway: Hermes with Telegram channel configuration
This repository is a public technical representation of the system. It does not claim that every reference script is an exact byte-for-byte export from the private vault; the scripts are sanitized, runnable equivalents of the operational contracts and checks used to maintain AustinOS.