The Karpathy Pattern

Background

In April 2026, Andrej Karpathy published a gist describing a pattern for building personal knowledge bases using LLMs. The core insight:

Most people’s experience with LLMs and documents looks like RAG: you upload a collection of files, the LLM retrieves relevant chunks at query time, and generates an answer. This works, but the LLM is rediscovering knowledge from scratch on every question. There’s no accumulation.

The solution: instead of per-query retrieval, have the LLM incrementally build and maintain a persistent wiki.

The Three-Layer Architecture

┌─────────────────────────────────────────────────────────┐
│  Layer 1: Raw Sources (Immutable)                       │
│                                                         │
│  Articles, papers, transcripts, images, data files.     │
│  The LLM reads these but NEVER modifies them.           │
│  Append-only — external changes create new snapshots.   │
└─────────────────────────────┬───────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────┐
│  Layer 2: The Wiki (LLM-Maintained)                     │
│                                                         │
│  Summaries, entity pages, concept pages, comparisons.   │
│  Cross-referenced, with contradictions flagged.         │
│  Human reads; LLM writes and maintains.                 │
└─────────────────────────────┬───────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────┐
│  Layer 3: Schema (Configuration)                        │
│                                                         │
│  Defines wiki structure, conventions, workflows.        │
│  Co-evolved over time as the wiki grows.                │
└─────────────────────────────────────────────────────────┘

Key Operations

Ingest

Add a source → LLM reads it, writes a summary page, updates the index, updates related existing pages. A single source may touch 10-15 wiki pages.

Query

Ask a question → LLM searches the wiki, reads relevant pages, synthesizes an answer with citations. Good answers can be filed back into the wiki.

Lint

Periodic health check for contradictions, stale claims, orphan pages, missing cross-references.

How Hermes Wiki Implements This

Karpathy Concept	Hermes Wiki Implementation
Raw sources (immutable)	raw/ directory with append-only snapshots, sha256 provenance, drift detection
LLM-maintained wiki pages	entities/, concepts/, comparisons/, sources/ — markdown with YAML frontmatter
Schema / CLAUDE.md	SCHEMA.md per wiki — domain contract, taxonomy, page thresholds, propagation rules
Index	index.md — sectioned page catalog with one-line summaries
Log	log.md — append-only attributed action record
Ingest operation	hermes-wiki ingest — pluggable classify → process → propagate → commit pipeline
Query operation	hermes-wiki search / wiki_search tool — BM25 FTS5 ranked results
Lint operation	hermes-wiki lint — 19 automated health checks
Knowledge compounds	Propagation rules update related pages on every ingest
Human curates sources	Humans drop sources in raw/inbox/; agents do the rest

What Hermes Wiki Adds Beyond the Pattern

Karpathy’s gist describes the conceptual pattern. Hermes Wiki makes it operational:

Capability	Description
Multi-wiki	Multiple domain-scoped wikis with independent schemas
Attribution	Every change tracked to agent, profile, human, or cron job
Privacy	Profile-scoped visibility with whitelist/blacklist
Agent integration	Tools injected into agent system prompts; agents use wikis in conversation
Dashboard	Web UI for browsing, searching, and managing wikis
Kanban linkage	Bidirectional wiki-page ↔ task references
Pluggable pipeline	Custom classifiers and processors with trust gating
Monitors	Automated source sweeps (arxiv, RSS, URLs) via cron
Projection versioning	SQLite rebuilds are versioned for triage
Standalone-first	Runs without Hermes installed; adapters wire into real deployments

Philosophy

The human’s job is to curate sources, direct the analysis, ask good questions, and think about what it all means. The LLM’s job is everything else.

— Andrej Karpathy

Hermes Wiki embodies this division of labor. Humans choose what to learn about (drop sources, ask questions). Agents handle the mechanical work of synthesis, cross-referencing, indexing, and maintenance.