LLM Wiki — Hermes Agent Skill

Build and maintain a persistent, compounding knowledge base as interlinked markdown files.

Based on Andrej Karpathy's LLM Wiki pattern, extended with health checks, knowledge graph visualization, and structured governance.

Unlike traditional RAG (which rediscovers knowledge from scratch per query), the wiki compiles knowledge once and keeps it current. Cross-references are already there. Contradictions have already been flagged.

Version: 2.2.0 | License: MIT

Update Summary (Human Control Layer)

• Added a Human Control Layer for large LLM-compiled wikis: overview.md, maps.md, questions.md, principles.md, decisions.md, plus queries/inbox.md for candidate questions.
• Added a mandatory return-flow gate after ingest/query/organizing: update the relevant human-owned entry page only when the big picture, topic map, long-term questions, principles, or concrete decisions change; otherwise record Human layer: no update in log.md.
• Clarified that overview.md is a living synthesis, not a forced dump: update it only when the overall picture changes, to keep the human layer readable and actually ownable.
• Added query-value governance: AI captures candidate questions when they meet value criteria; the human calibrates/promotes them instead of letting the LLM decide final knowledge automatically.
• Preserved existing operational improvements: reverse-chronological log.md, source-vs-raw separation, health/lint helpers, graph tooling, and post-ingest validation.

What's Inside

SKILL.md

The core behavioral specification for the agent. Covers:

• Directory structure — three-layer architecture plus Human Control Layer
• Ingest workflow — raw faithful capture, wiki compilation, return-flow gate, post-ingest validation, and domain-specific templates (diary, meeting notes)
• Query workflow — synthesize answers from compiled knowledge with citations, capture valuable questions into queries/inbox.md, and update human-owned pages when judgment changes
• Health check — zero-LLM pre-flight structural integrity check
• Knowledge graph — two-pass build (deterministic wikilinks + semantic inference) with vis.js visualization
• Lint — automated checks including graph-aware analysis (phantom hubs, hub stubs, fragile bridges)
• Governance — Human Control Layer, tag taxonomy, contradiction policy, archive/rotation lifecycle

references/

Standalone Python tools:

Script	Purpose
health-check.py	Structural health checks (empty files, index sync, log coverage) — zero LLM calls
build-graph.py	Knowledge graph builder with NetworkX + Louvain community detection + vis.js HTML output
lint-scripts.py	Automated lint functions (orphans, broken links, frontmatter, tags, page size, reverse-log rotation check/execute)

All scripts read WIKI_PATH from environment (defaults to ~/Desktop/wiki).

Install as Hermes Skill

# If using hermes-skills-backup:
cd ~/.hermes/skills/research
git clone https://github.com/Robs87/llm-wiki.git llm-wiki

Quick Start

1. Set WIKI_PATH in your ~/.hermes/.env
2. Tell the agent to initialize a wiki for your domain
3. Start ingesting sources — the agent handles the rest

"Create a wiki for AI/ML research"
"Ingest raw/papers/attention-is-all-you-need.md"
"What does the wiki say about transformer architecture?"
"Check wiki health"
"Build the knowledge graph"

Comparison with SamurAIGPT/llm-wiki-agent

This skill incorporates the best ideas from that project (health checks, knowledge graph, post-ingest validation, phantom hub detection) while preserving stronger governance (SCHEMA.md with tag taxonomy, contradiction handling policy, archive/rotation lifecycle, session orientation protocol).

Credits

• Andrej Karpathy — original wiki pattern
• SamurAIGPT/llm-wiki-agent — health check, graph, and validation inspiration