LLM Wiki Coordination

A multi-agent governance layer for self-maintaining LLM wikis.

If you already run a Karpathy-style LLM wiki and use more than one AI agent (Claude, Codex, Gemini, or any combination), this layer adds the missing protocol for them to work together coherently across sessions.

What this gives you

• Consensus protocol — formal agent-level endorsement of structural documents, with stale invalidation, soft canonical, and explicit revision tracking.
• Async dialogue threads — agents exchange Q&A across separate sessions, with statuses (open → pending-<agent> → crystallizing → crystallized) that promote the result into a wiki workflow.
• Honest back-channel — a separate folder where each agent records observations without diplomatic filter, kept distinct from the canonical wiki voice.
• Boundary handling — explicit playbook + redirect algorithm for handling constrained requests without dead-ending the user.
• Shared user profile — a single user-working-profile.md that all agents read, replacing drifting per-agent CLAUDE.md / CODEX.md / GEMINI.md duplicates.
• Memory tiers — frontmatter tier: field (working / episodic / semantic / procedural) so agents read in priority order and skip episodic logs at startup.
• Typed relations — graph edges between entities (feeds_into, depends_on, opposes, validates, can_distort_into, …) instead of flat wikilinks.

Why this exists

The Karpathy LLM wiki pattern works beautifully for one agent. With multiple agents and sessions weeks apart you get:

• Three CLAUDE.md / CODEX.md / GEMINI.md files drifting apart
• Each agent unaware of what the others have done
• No formal way to mark a document "endorsed by all"
• Refusals that derail the user instead of finding the closest valid route
• No place to record honest in-session observations without breaking the canonical voice
• Open questions between agents with no asynchronous channel to resolve them

This repo provides a drop-in wiki/agents/ folder structure plus a few additions to your existing AGENTS.md. Nothing in your existing wiki has to be rewritten.

Quick install

# In your existing LLM wiki repo
cp -r path/to/llm-wiki-coordination/templates/agents wiki/agents
cp path/to/llm-wiki-coordination/templates/workflows/* wiki/workflows/
cp path/to/llm-wiki-coordination/templates/concepts/* wiki/concepts/

Then merge the protocol additions in docs/installation.md into your AGENTS.md.

See docs/installation.md for the full guide.

Repo layout

docs/                      Specifications and the "why"
  overview.md
  installation.md
  protocol/                Each protocol explained
  concepts/                Memory tiers, typed relations
templates/                 Drop-in files for your existing wiki
  agents/
  workflows/
  concepts/
examples/                  Before/after, sample threads

Lineage

This layer stands on three earlier ideas:

• Andrej Karpathy's LLM Wiki gist — the immutable raw/ plus LLM-maintained wiki/ pattern.
• rohitg00's LLM Wiki v2 gist — memory tiers, typed relations, and the idea of multi-agent mesh sync (proposed at gist level; this repo implements it).
• A working private second-brain implementation (April 2026) where these protocols were developed across actual Claude / Codex / Gemini collaboration before being extracted into this repo.

Status

v0.1.0 — initial extraction. Protocol is in active use in the source implementation but interfaces may evolve.

License

MIT