Skip to content
LlmWikis.orgLLM Wiki handbook
UAIX canonical
Menu
UAIX canonical
Source map LLM Wiki handbook llms.txt What is an LLM Wiki? Setup wizard Checklist Specification Metadata Navigation vs RAG vs AI Memory Source policy UAI-1 spec

LlmWikis knowledge page

Start

Start Here gives first-time builders a practical route through the LLM Wiki pattern. The goal is not to upload every file you own. The goal is to create a small, reviewable system where raw evidence stays preserved, compiled pages improve over time, and the next agent can understand what changed without reading a private chat transcript.

What to understand first

An LLM Wiki is a stateful knowledge base, not a one-off chat answer. The agent reads immutable sources, writes durable markdown pages, updates an index and log, and follows a compact root instruction file so each ingest, query, and lint pass leaves the graph easier to use.

First concept

What is an LLM Wiki?

Understand the difference between a governed LLM Wiki, a normal wiki, AI Memory, and RAG.

First workflow

Ingest a source

Turn one source into source summaries, concept updates, links, contradictions, and a log entry.

Starter vocabulary

Term Plain meaning Why it matters here
Raw source An original file, transcript, page, paper, dataset, or note preserved without AI edits. It is the evidence layer that lets a reviewer rebuild or challenge the wiki.
Compiled wiki The markdown graph the agent may update after review. It stores the useful synthesis so future sessions do not start from zero.
Hot context The small file set an agent loads by default before acting. It should contain current truth, constraints, next work, and checks, not every historical note.
Cold memory Long-lived source history, old research, archived handoff snapshots, and recovery evidence. It stays searchable and recoverable without flooding every agent startup.
Schema file A root instruction file such as AGENTS.md or CLAUDE.md. It defines write permissions, page shape, source rules, and done checks.
index.md The routing catalog for every useful page. The agent can choose a small page set instead of scanning everything.
log.md An append-only record of ingest, query, lint, review, and correction events. The next agent can see what changed, why, and which sources were involved.
Lint A maintenance pass for links, stale claims, contradictions, missing sources, and orphan pages. It keeps fluent but unsupported text from becoming trusted memory.

Learning paths

Reader Sequence Exit condition
Personal builder Definition -> build guide -> starter template -> three layers -> ingest -> personal example. You can add a source and ask a question without reading raw files again.
Team maintainer Definition -> structure -> trust model -> metadata -> governance -> security/privacy -> checklist. You have review gates, source boundaries, and a safe contribution path.
Technical writer Content types -> metadata standard -> navigation -> templates -> examples -> llms.txt source map. Your docs can be read by humans and routed by AI crawlers.

First 30 minutes

  1. Pick one bounded topic. A project, research thread, support area, or codebase module is better than an entire life archive.
  2. Set a context budget. Decide which files are hot startup context and which belong in cold memory with a pointer.
  3. Install the starter structure. Use the starter bundle or create the same files manually.
  4. Create three layers. Make source evidence, compiled pages, and a compact root instruction file before the first ingest.
  5. Add one source. Start with a source you can inspect manually so you can judge whether the compiled page is faithful.
  6. Write index and log first. Even a tiny wiki should route itself and record what happened.
  7. Run a lint pass. Check source links, contradictions, page status, and whether the next reader has a clear path.
Good first success: one source summary, one concept page, one index entry, one log entry, and one visible unresolved question. That is enough to prove the loop.