MyForge Vault 11.11#

An open-source 8-axis methodology + working tooling for evolving a personal Obsidian-vault into a self-improving knowledge-system. Made by MyForge Labs. Augmented intelligence — NOT AGI, NOT hype. Hungarian+English docs, MIT.

📚 Docs site · 🎬 3-min demo · 🇭🇺 Magyar · 📋 FAQ · 🗺️ Architecture · 📖 The build story (3,900 words)

What is this#

MyForge Vault 11.11 (internal codename: SV) is an 8-axis architecture + 85+ production scripts + 274 evergreen wiki pages + 46 ADRs + 126 audits that turns a classic Obsidian-vault into a self-improving knowledge-system shared by three CLI AI agents (Claude Code, Codex, Gemini). Measurable numeric results, clear OSS scope, MIT-licensed, $0 marginal cost.

If you have 90 seconds: read the FAQ. If you have 5 minutes: read the architecture overview. If you have 15 minutes: read what I learned building it in 5 hours.

The 11.11 in the name carries two meanings: - 🏢 MyForge Labs founding signal — the company's 11.11@myforgelabs.com email predates this vault - 🔧 Session-orchestration primitive — every workflow runs through the 11.11* CLI family (11.11start, 11.11stop, 11.11note, 11.11focus, 11.11ls, 11.11crystallize, 11.11worker) — the connective tissue that makes the 8 axes work as one system

The methodology starts from Karpathy's LLM-Wiki pattern: the "raw input" (10-raw) → "distilled knowledge" (11-wiki) crystallization workflow. SV extends this through evolution along 8 independently developable axes.

The 8 axes#

The 7 most important artifacts (NotebookLM-recommended)#

1. Subagent-fanout dispatcher — 174× parallel LLM-task, $0 cost (Claude Code subscription)
2. load-session-context — MemGPT-style virtual context loader, 75% token savings
3. vault-search-server — Unix-socket daemon, 80× speedup (14s→165ms) + Memgraph 280× speedup
4. Bias-mitigated G-Eval — Claude-to-Claude self-enhancement debiasing, 96.7% calibration agreement
5. Smart-trigger NLI cascade — fast-baseline → expensive-only-if-needed, 5-10× cost-savings
6. 4-layer Safety-Gate — ENV + script + git-hook + Critic review (RSI guardrail)
7. Sprint Day-0 Skeleton-first — ~5× faster Week 1 implementation

Measured results (2026-04-23 → 2026-05-19, 27 days)#

Compare to other memory + agent-OSS projects#

This is an honest, opinionated map of where SV sits. Each project below is excellent at what it does; SV is a composite built around a different shape (markdown-vault first, three-CLI-agent-shared, local-by-default).

Where SV is NOT the right pick:

• If you want a hosted memory SaaS with multi-tenant isolation and a Python SDK, use mem0
• If you want a persistent agent runtime with full state checkpointing, use Letta
• If you want GraphRAG specifically (Microsoft's community-detection + hierarchical summarization), use the original GraphRAG
• If you want a simple key-value memory with confidence scoring and zero graph infrastructure, use agentmemory

Where SV IS the right pick (after evaluating the above):

• you run multiple CLI agents on one machine
• you already use Obsidian and the markdown-first format matters
• you want the build-as-you-go vault to BE the artifact, not a side-effect
• you want a reference implementation of Karpathy's LLM-Wiki pattern that actually runs

Compare to other agent-skill repos#

Different category — Pocock/skills, obra/superpowers, tinyhumansai/openhuman are skill-libraries that any agent loads. SV shares this property (962 indexed SkillChunks) but doesn't compete on the skill-library axis. Use these alongside SV:

Quick start (≈ 15 min, all-deps fresh)#

# 1. Clone
git clone https://github.com/MyForgeLabs/myforge-vault-1111.git
cd myforge-vault-1111

# 2. Memgraph CE (Docker)
docker run -d --name memgraph -p 7687:7687 memgraph/memgraph:latest

# 3. Python deps
make install          # or: pip install -r requirements.txt

# 4. Embed the wiki content
./scripts/vault-embed.py --backfill 11-wiki/

# 5. Try a search
./scripts/vault-search "Karpathy LLM-Wiki pattern"

Verify:

make test             # runs the LongMemEval-S fast regression-gate
make build-docs       # mkdocs build --strict

make help lists everything. See the FAQ for the "works on my machine" checklist (OS support, Python version, common friction).

Architecture in one diagram#

A full Mermaid diagram lives in 11-wiki/architecture-overview.en.md. The short version:

   📥 INPUT          🔮 CRYSTALLIZE        🧠 MEMORY          ✨ DISTILLED
   sessions    ──▶   /11.11stop hook  ──▶  KO-DB (SQLite) ──▶ 11-wiki/
   raw/external      G-Eval scorer         Memgraph CE        07-Decisions/
   browser-hist      Constitutional       (vector + graph)    06-Audits/
   3 CLI agents      Critic gate (4-lyr)   bge-m3 + reranker  02-Projects/
                          │                      │
                          ▼                      ▼
                     🛠️ TOOLING (vault-search · vault-mcp · 962 skills)
                          │
                          ▼
                     📊 EVAL + RSI (LongMemEval-S gate · GEPA · Tier-2 RSI)
                          │
                          ▼
                     🎙️ COGNITIVE (NotebookLM cross-project synthesis)

Reproducibility#

The full methodology is architecture-level reproducible through the 07-Decisions/ ADRs + 11-wiki/ evergreen wikis. Every script is idempotent, ENV-flag-gated, default-OFF safety pattern.

Positioning (transparent)#

MyForge Vault 11.11 is NOT a "Pocock-skills alternative" or "openhuman challenger". The methodology is an 8-axis composite architecture with measurable results, used on MyForge Labs' own Obsidian-vault, published as open source. Goal: industry peer feedback + anyone else reproduces it on their own vault.

Who's behind it#

MyForge Labs — small Hungarian engineering shop building agent-skill infrastructure, multilingual web platforms, and AI-augmented operational tooling. Founded around 11.11.

Maintainer: Peti Markovics (@petimarkovics · peti.markovics@gmail.com).

Contributors#

This project is AI-aided by design, not by accident. The three CLI agents listed below are co-collaborators, not tools. Every commit is stamped with an AGENT= env-var so you can git log --grep='AGENT=' to see which agent did which work.

If you want to be listed here, open a PR. Human OR agent contributors welcome.

Acknowledgements#

• Andrej Karpathy — for the original LLM-Wiki gist this whole project is built on
• Memgraph team — for shipping a CE with native vector-index and no licensing wall
• BAAI — for bge-m3 and bge-reranker-v2-m3 (multilingual, CC-BY)
• Anthropic / OpenAI / Google — for the CLI agents that made the AI-aided-build feasible
• The Obsidian community — for normalizing markdown-first knowledge work

License#

MIT — see LICENSE. Cherry-pick freely, attribution-friendly.

Related#

• Architecture decision records (46)
• Evergreen wikis (274)
• Audits (126)
• FAQ
• Architecture overview (with Mermaid diagram)
• Hungarian README
• The Karpathy-style build story (3,909 words)

@software{markovics_2026_myforge_vault,
  author       = {Markovics, Peti},
  title        = {{myforge-vault-1111: A self-improving Obsidian
                   vault for CLI AI agents}},
  year         = 2026,
  version      = {1.0.9},
  url          = {https://github.com/MyForgeLabs/myforge-vault-1111}
}