Kihagyás

FAQ — MyForge Vault 11.11#

The questions you'll have before / during / after reading the rest of the project. Curated from the kinds of things people ask new OSS launches.

"What is this, in one sentence?"#

A locally-hosted, self-improving markdown knowledge-vault that three CLI AI agents (Claude Code, Codex, Gemini) share — built around Karpathy's LLM-Wiki pattern with a Memgraph knowledge-graph, G-Eval auto-crystallization, and a 4-layer Constitutional safety-gate.

"Should I use this?"#

Use it if you:

  • run multiple CLI AI agents on the same machine and want them to share state
  • already use Obsidian and feel its native graph is "shallow"
  • want a working reference implementation of Karpathy's LLM-Wiki pattern rather than a blog post
  • are a researcher / agent-builder who needs an instrumented memory testbed
  • want to fork a working agentic-OS and adapt it to your domain

Do NOT use it if you:

  • want a SaaS — this is local-only, by design
  • need a hosted vector DB — this uses Memgraph + bge-m3 locally
  • are looking for a Cursor/Cline plugin — this is a CLI/agent-stack
  • expect first-class Windows support — it runs on Linux + macOS; Windows is best-effort via WSL

"Is this AGI?"#

No. It's augmented intelligence — three CLI agents sharing a structured notebook that gets better over time. The agents do the work; the vault remembers. No claims of recursive self-improvement reaching superhuman levels; the RSI loop is gated behind 4 explicit safety layers and --apply mode is locked by default. See the RSI safety gate.

"How much does it cost to run?"#

$0 marginal cost if you have a Claude Code subscription (and / or a NotebookLM subscription for podcast generation).

The numeric details:

  • Claude Code subscription ($20/mo) — covers all agent work including bulk subagent-fanout LLM mutations (10K+ ops across the 26-day build)
  • NotebookLM (free tier or $20/mo Plus) — used for cross-project synthesis and 2-host podcast generation
  • Local infra — Memgraph CE (free), bge-m3 (open-weights, free), Python stack (free), a VPS with ~32 GB RAM + 8 cores recommended

If you skip the Claude Code subscription and use the Anthropic API directly, add ~$80/mo for an equivalent workload. The subagent-fanout pattern specifically exists to avoid this.

"How do I install it?"#

Five steps, ~15 minutes if all the deps are pre-installed:

# 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"

See README.md for the full quickstart with verification steps.

"Why three CLI agents instead of one?"#

Different agents have different strengths and pricing models, and each makes mistakes that the others catch:

  • Claude Code — best at long-form code generation, has the Task tool for the subagent-fanout pattern; primary workhorse
  • Codex — strong at code refactoring + IDE-integrated workflows; second opinion on architecture
  • Gemini — strong at multimodal + has unique session-context patterns; third perspective

They share a single AGENTS.md system prompt (symlinked across all three CLI homes), the same vault, the same 11.11* session-orchestration scripts. The AGENT= env-var stamps every commit so you can git log --author to see which agent did what.

"What's in the vault by default?"#

A complete working example with my own content:

  • 258 evergreen wikis (Hungarian primary, 71 English translations)
  • 45 ADRs (Architecture Decision Records)
  • 104 audits (one-shot reports)
  • ~80 session-logs (some scrubbed for public)
  • 13,800+ KO-DB triplets extracted from the above
  • 8,997 typed entities in Memgraph (with 24,606 edges)
  • 962 indexed skills (SKILL.md files)
  • 3 NotebookLM 2-host podcast episodes

You can use this as-is to study the architecture, or rm -rf 11-wiki/* 07-Decisions/* 06-Audits/* 08-Sessions/* and put your own content in.

"What's the 'agentic OS' framing?"#

Not marketing — a literal description of how it works.

The three CLI agents are processes that read/write a shared filesystem (the vault). The 11.11* family is the IPC mechanism (session-start, note-append, session-stop). The vault-search daemon is a long-running service (like a kernel module). Memgraph is the database (like a filesystem index). G-Eval is the access-control gate. Constitutional AI Tier-2 is the privileged-mode escalation path (locked).

An OS by analogy. NOT a desktop OS, NOT a kernel — but the same shape.

"Why Memgraph and not Neo4j / NetworkX / DuckDB / SQLite?"#

  • vs Neo4j: Memgraph CE 3.9 native vector-index (no licensing wall), 280× speedup over numpy-cosine, p95 2.6 ms. Neo4j Community is roughly equivalent for graph queries but the vector index lives behind a paid tier in their hosted product.
  • vs NetworkX: no persistence, no concurrent access, no vector index
  • vs DuckDB: great for analytical SQL, weaker for graph traversals (Cypher in DuckDB is third-party + immature as of 2026-05)
  • vs SQLite (KO-DB): we already use SQLite for the triplet store! The KO-DB and Memgraph are complementary, not alternatives — SQLite handles the substring + filter top-k retrieval (60 ms), Memgraph handles the graph traversal + vector search

See memgraph-ce-feature-limits.en.md for the unrosy details.

"Is the AI-aided-build disclosed?"#

Yes — explicitly. The repo has 70+ commits in the launch sprint alone, most of them with AI-agent co-authorship. The AGENT= commit-trailer is by design. The Karpathy-style essay (What I learned...) is candid about which parts were AI-aided.

This isn't hidden, it's the point: a working artifact built BY AI agents USING the vault we're publishing. The vault is what makes the AI-aided build sustainable.

"Can I run this without exposing my vault content?"#

Yes. Default install ingests only the files you point it at. No telemetry, no analytics, no remote endpoints. Memgraph is local. bge-m3 is local. The only outbound call is when you ask the agent to fetch something explicitly (firecrawl, NotebookLM, etc.).

The opt-in browser-history bridge has a --dry-run default; VAULT_BROWSER_INGEST_APPLY=1 is required to actually write to the vault.

"Is there a Discord / Slack / Matrix?"#

Not yet. Use GitHub Discussions for now. A real-time chat will appear if there's demand and a community to support it (otherwise it's a graveyard).

"Can I contribute?"#

Yes. Three flavors:

  1. Share your vault pattern — open a [pattern] issue or a Discussion
  2. Add a wiki — fork, write under 11-wiki/<your-page>.md, PR
  3. Fix code — open an issue first if >50 LOC; PR with tests if there's a test framework for that subsystem

See CONTRIBUTING.md. The AI agents are co-collaborators on this project — your PR will probably be reviewed by both a human and a Claude Code agent. That's fine; both make mistakes the other catches.

"Hungarian-first feels weird"#

It's not weird, it's honest. I'm Hungarian, the original session-logs are in Hungarian, and bge-m3 handles HU↔EN semantic search well enough that the hybrid is genuinely a feature (not a bug). 71 of the 258 wikis (28%) are translated to English; the most-trafficked ones are translated first, the rest as community demand arises.

If you want to help translate, see the i18n: <page> PR pattern.

"What's NOT in this repo (yet)?"#

The next-session backlog (deferred from 2026-05-19):

  • Temporal-KG SCD2 layer — valid_from/valid_until on KO-DB facts for time-travel queries
  • Transaction-aware atomic_append_jsonl — for the 2 vault-ko-remap-legacy sites that need it
  • Real-LLM Critic for the sleep-consolidation pipeline (currently a rule-based mock + a pending-file interface)
  • Cloudflared tunnel for the vault-mcp STDIO server to be reachable from claude.ai web/mobile
  • CI workflows — partially landed in v1.0.1; full coverage in v1.0.2

See the repo improvement audit for the post-launch roadmap.

"Anything else?"#

Star the repo. Open a Discussion if anything here is unclear. The single most-valuable thing you can do is share your vault pattern, even if it's totally different.