Karpathy LLM Memory Wiki

Compilation-over-retrieval knowledge wiki for OpenClaw agents. Drop sources in, get structured cross-referenced pages out. Knowledge compounds instead of disappearing. Based on Andrej Karpathy's LLM Wiki pattern.

Wiki System

"The tedious part of maintaining a knowledge base is not the reading or thinking — it's the bookkeeping." — Andrej Karpathy

Most AI memory systems retrieve and forget. You ask a question, the agent fetches context, answers, and the insight vanishes into chat history. Next session, same question, same fetch, same synthesis from scratch.

This skill implements Karpathy's compilation-over-retrieval pattern (original gist): instead of fetching raw sources every time, the agent builds and maintains a structured wiki — cross-referenced markdown pages that compound knowledge over time. The human curates what goes in; the agent handles the summarization, cross-referencing, and maintenance that humans typically abandon because the overhead grows unbearably.

Three Layers

1. Raw Sources (immutable) — PDFs, articles, papers, URLs. You drop them in, the agent reads but never modifies them.
2. The Wiki (LLM-maintained) — structured markdown pages in memory/wiki/ with frontmatter, claims, cross-references. The agent creates, updates, and maintains these automatically.
3. The Schema (WIKI-SCHEMA.md) — rules for structure, naming, and workflows. Co-evolves with your knowledge base.

Three Operations

• Ingest — source goes in, agent extracts claims, creates/updates wiki pages, adds [[cross-references]] across related pages. A single source can touch multiple existing pages.
• Query — agent checks the wiki first, synthesizes answers with citations, and optionally files gaps back to the ingestion queue. Insights compound instead of evaporating.
• Lint — weekly health check: stale pages, broken links, orphan pages, contradictions, missing metadata. The wiki stays clean as it grows.

Compatibility

Works out of the box with

• OpenClaw >= 2026.1.29 (only hard requirement)
• Builtin memory backend (default) — the setup script adds memory/wiki to memorySearch.extraPaths so wiki pages are searchable via memory_search. No manual config needed.
• Dreaming (bundled, opt-in) — wiki and dreaming coexist cleanly. Dreaming consolidates daily conversation signals into MEMORY.md; the wiki compiles external knowledge into memory/wiki/. Different inputs, different outputs, never touching each other's files. Dreaming can even queue wiki candidates to the ingestion queue — the 4AM ingest cron picks them up 90 minutes after dreaming finishes at 2:30AM.

Enhanced with (optional)

• QMD — local search sidecar with reranking and query expansion. Recommended for wikis over ~50 pages. QMD indexes memory/**/*.md recursively, so wiki pages are auto-indexed every 10 minutes with no extra config.
• SuperMemory — cloud memory persistence. Adding memory/wiki/index.md to the sync script lets cross-session recall know what wiki topics exist, even on different devices.
• Embedding provider (OpenAI, Gemini, Voyage, Mistral) — enables semantic/vector search alongside keyword matching. Especially useful when wiki pages use different terminology than your queries.
• Honcho — cross-session semantic search with user modeling. Wiki pages indexed automatically.
• Dual-memory plugin (memory-core + SuperMemory) — if you run both backends via a composite plugin, wiki search results merge with cloud memories automatically. Best of both worlds.
• Obsidian — the [[wikilink]] cross-references are Obsidian-compatible. Point Obsidian at memory/wiki/ to browse, visualize the knowledge graph, and follow links.

What the wiki agent never modifies

• DREAMS.md — owned by Dreaming diary
• memory/YYYY-MM-DD.md — owned by daily compaction
• memory/dreaming/ — owned by Dreaming reports
• memory/projects/active/ — owned by project workflows
• memory/decisions.log, memory/learnings.log — owned by tracking systems
• Your existing memory files, projects, logs — all untouched

Note: The setup script appends a wiki reference section to MEMORY.md once during installation (so the agent knows the wiki exists). After setup, the wiki system never modifies MEMORY.md again — Dreaming retains full ownership.

How It Works

text

You drop sources into work/wiki-sources/
        |
        v
[Wiki Ingest - cron or manual]
  Reads sources, extracts claims, creates/updates pages
        |
        v
memory/wiki/*.md (structured pages with frontmatter + cross-references)
        |
        v
[OpenClaw Memory Index]
  memory_search finds wiki pages alongside your other memory files

Wiki pages live inside memory/ so every OpenClaw memory backend indexes them automatically. No extra configuration needed.

Installation

Basic setup (directories, templates, MEMORY.md reference):

bash

bash skills/karpathy-wiki/scripts/wiki-setup.sh

Full setup with automated cron jobs and SuperMemory cloud sync:

bash

bash skills/karpathy-wiki/scripts/wiki-setup.sh --all --tz America/New_York

Flags:

• --tz TIMEZONE — timezone for cron jobs (default: UTC)
• --with-cron — create ingest (daily 4AM) and lint (Sunday 3:30AM) cron jobs
• --with-sync — add wiki index to SuperMemory sync script (enables cloud persistence)
• --all — enable both --with-cron and --with-sync

The basic setup creates:

1. memory/wiki/ with index.md and log.md
2. work/wiki-sources/ with pdfs/, articles/, papers/ subdirs
3. WIKI-SCHEMA.md at workspace root
4. Ingestion queue at work/wiki-sources/ingestion-queue.md
5. Wiki reference section appended to MEMORY.md (if present)

Cron jobs and SuperMemory sync are opt-in — you control what gets automated.

Uninstall

bash

openclaw cron delete <ingest-job-id>    # Get IDs: openclaw cron list | grep Wiki
openclaw cron delete <lint-job-id>
rm -rf memory/wiki/ work/wiki-sources/ WIKI-SCHEMA.md

Usage

Adding knowledge

During a conversation:

1. Save the source to work/wiki-sources/
2. Add it to work/wiki-sources/ingestion-queue.md:

   markdown

   ## Pending
   - [ ] articles/my-article.md | Added 2026-04-07
   - [ ] https://example.com/paper | Added 2026-04-07
   

3. Say "wiki ingest" — the agent processes the queue immediately

Trigger ingest outside a conversation:

bash

openclaw cron run <ingest-job-id>

Automated: Cron processes up to 3 sources daily at 4AM.

Querying the wiki

The agent checks memory/wiki/index.md for relevant pages, reads them, and synthesizes answers with citations. If a question reveals a knowledge gap, it adds an entry to the ingestion queue for future research.

Manual lint

Say "wiki lint" to run a health check anytime.

Page Format

Pages use domain-prefixed kebab-case names with YAML frontmatter:

yaml

---
title: "Senolytics Overview"
domain: longevity
created: 2026-04-07
updated: 2026-04-07
sources:
  - work/wiki-sources/papers/unity-trial.pdf
tags: [senescence, aging]
confidence: high
status: active
---

# Senolytics Overview

## Summary
2-3 sentence overview.

## Key Claims
1. Claim with source attribution

## Open Questions
- Unresolved questions

## Related
- [[longevity-cellular-reprogramming]]

Domains: bio, vita, longevity, tech, ops, finance, people (customizable in WIKI-SCHEMA.md)

See skills/karpathy-wiki/templates/page-template.md for a starter.

File Layout

text

WIKI-SCHEMA.md                         <- Rules and conventions
memory/wiki/index.md                   <- Page catalog (<4000 tokens)
memory/wiki/log.md                     <- Activity log (append-only)
memory/wiki/{domain}-{topic}.md        <- Wiki pages
work/wiki-sources/                     <- Raw source documents
work/wiki-sources/ingestion-queue.md   <- Processing queue
skills/karpathy-wiki/scripts/            <- Ingest and lint prompts
skills/karpathy-wiki/templates/          <- Scaffold templates
skills/karpathy-wiki/references/         <- Background on the pattern

Cron Schedule

Job	Schedule	Recommended Model	Purpose
Wiki Ingest	4:00 AM daily	Reasoning-capable (e.g., Claude Sonnet, GPT-4o, Gemini Pro)	Process ingestion queue — needs synthesis and cross-referencing
Wiki Lint	Sunday 3:30 AM	Fast/cheap (e.g., Claude Haiku, GPT-4o-mini, Gemini Flash)	Health check — mechanical validation, no creative thinking needed

Timezone configurable via --tz during setup. Ingest runs 90 minutes after Dreaming (2:30AM) to avoid conflicts. Model is set during cron creation — change anytime with openclaw cron edit <id> --model <model>.

Troubleshooting

• Pages not searchable: openclaw memory status to check index. Try openclaw memory index --force.
• Cron not firing: openclaw cron list for status. Restart gateway if needed.
• Index too large: At 80+ pages, schema supports domain sub-indices. See WIKI-SCHEMA.md.
• Dreaming conflicts: Impossible by design. Wiki and Dreaming write to different files with a 90-minute scheduling buffer.

Credits

Inspired by Andrej Karpathy's LLM Wiki pattern. Implementation adapted for OpenClaw's memory architecture with compatibility across all memory backends (builtin, QMD, Honcho), optional Dreaming integration, and SuperMemory cloud persistence.

See skills/karpathy-wiki/references/karpathy-llm-wiki.md for the full concept breakdown.