Hermes Agent Ships a Bundled LLM Wiki Skill Based on Karpathy's Knowledge Base Pattern

On June 15, YanXbt published a thread breaking down the LLM Wiki skill that ships bundled with Hermes Agent. The thread collected 80 likes and 168 bookmarks in its first 12 hours. It is the third in a series of deep-dive posts YanXbt has published about Hermes Agent internals, following a breakdown of the eight internal loops that drive agent behavior and a guide to the 9-hour overnight autonomous workflow.

The skill implements Andrej Karpathy's LLM Wiki pattern - a persistent, compounding knowledge base built from interlinked markdown files. Unlike retrieval-augmented generation (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. Every new ingest compounds on prior work.

The bundled skill (v2.1.0, shipping at skills/research/llm-wiki/SKILL.md) was originally merged in PR #5635 in April alongside the Skill Config Interface. It was upgraded from v1.0 (274 lines) to v2.0 (400 lines) after live testing revealed gaps in session continuity and page management, then to v2.1.0 with headless Obsidian sync support.

Three-Layer Architecture

The wiki is a directory of markdown files with a three-layer design. Layer 1 holds immutable raw sources. Layer 2 holds agent-owned wiki pages. Layer 3 is the governing schema.

Every wiki page carries YAML frontmatter with required fields: title, created, updated, type, tags, and sources. Optional quality signals include confidence (high/medium/low) and contested (boolean) for flagging unresolved contradictions. Raw source files also carry frontmatter with source_url, ingested date, and a sha256 hash of the body - enabling drift detection on re-ingest.

Three Core Operations

The skill defines three primary operation modes.

Ingest

When a user provides a source - URL, PDF, or pasted text - the agent executes a six-step pipeline:

1. Capture the raw source with frontmatter and SHA-256 hash
2. Discuss takeaways with the user (skipped in automated contexts)
3. Check what already exists by searching index.md and scanning existing pages for mentioned entities
4. Create or update wiki pages, with cross-references via [[wikilinks]], provenance markers (^[raw/articles/source.md]), and confidence annotations
5. Update navigation - new pages added to index.md, log entry appended
6. Report every file created or updated

A single source can trigger updates across 5-15 wiki pages. This compounding effect is the core value proposition - each ingest makes the next one cheaper, because more entities already have pages and more relationships are already mapped.

The skill enforces page thresholds: create a page only when an entity or concept appears in two or more sources, or is central to one source. Passing mentions do not warrant pages. When a page exceeds 200 lines, it is split into sub-topics. When content is fully superseded, the page moves to _archive/ with backlink cleanup.

Query

Querying follows a structured path: read index.md to identify relevant pages, run search_files across all .md content for key terms (necessary for wikis above 100 pages), read the matched pages, then synthesize an answer with citations.

Answers that represent substantial synthesis are filed back into the wiki as queries/ or comparisons/ pages. Trivial lookups are not filed - only results that would be painful to re-derive.

Lint

The lint command runs a 13-point health check across the entire wiki. It surfaces orphan pages with no inbound wikilinks, broken links pointing to nonexistent pages, index completeness gaps, frontmatter validation errors, stale content (pages whose updated date is more than 90 days behind the most recent source), contradictory claims, confidence signals, source drift via SHA-256 recomputation, oversized pages, tag taxonomy violations, and log rotation status.

Findings are grouped by severity: broken links, orphans, source drift, contested pages, stale content, and style issues.

Obsidian-Native

The wiki directory works as an Obsidian vault without any conversion step. [[wikilinks]] render as clickable links, Graph View visualizes the knowledge network, and YAML frontmatter powers Dataview queries. If the Obsidian skill is also enabled, setting OBSIDIAN_VAULT_PATH to the same directory lets both skills operate on the same vault.

For headless machines, the skill documents an obsidian-headless setup via systemd that syncs the vault continuously. An agent writes to ~/wiki on a server while the user browses the same vault in Obsidian on a laptop or phone - changes appear within seconds.

Session Continuity

The orientation protocol is the critical piece that prevents wiki degradation across sessions. Every session starts with three mandatory reads: SCHEMA.md for domain conventions and tag taxonomy, index.md for the full page catalog with one-line summaries, and the last 20-30 entries of log.md for recent activity.

Skipping orientation leads to duplicate pages, missed cross-references, and contradictions with existing content. The protocol is cheap - three file reads - and the skill marks it as "CRITICAL" in all-caps at the top of its operating instructions.

"Hermes Agent ships with a bundled skill for Andrej Karpathy's LLM Wiki pattern," YanXbt wrote in the thread. "A self-improving knowledge base that grows every time you feed it."

The thread connects the wiki skill to YanXbt's earlier overnight workflow article, where the wiki serves as the persistent memory layer that autonomous overnight sessions read from and write to - each night's run compounds on the previous night's output.

The LLM Wiki skill is one of several knowledge-management skills bundled with Hermes Agent, alongside the Obsidian skill (read, search, create notes in a vault) and arxiv integration. It occupies the research category in the skills catalog and supports Linux, macOS, and Windows.

[^1]: YanXbt. "Hermes Agent ships with a bundled skill for Andrej Karpathy's LLM Wiki pattern." X. June 15, 2026.

[^2]: Nous Research. "feat(skills): add skill config interface + llm-wiki skill" GitHub. Merged April 6, 2026.

[^3]: Andrej Karpathy. "LLM Wiki" GitHub Gist.