An LLM Wiki works because it refuses to blur storage, synthesis, and governance. Raw sources remain immutable. The wiki layer changes. The schema file tells the agent how to change it.
Architecture invariants
| Invariant | Design rule | Why it protects quality |
|---|---|---|
| Evidence is separate from synthesis | Keep originals in raw/ and compiled pages in wiki/. |
Bad summaries can be corrected without losing the source trail. |
| Navigation is explicit | Maintain index.md, log.md, page frontmatter, and typed links. |
The next human or AI can route by records instead of guessing from memory. |
| Claims have lifecycle state | Use draft, reviewed, source-needed, contradicted, stale, and archived labels. | Polished prose does not accidentally become trusted evidence. |
| Automation is staged | Analyze first, stage changes, review risky claims, then write and lint. | The wiki compounds without letting one model pass overwrite the graph blindly. |
Three-layer architecture
Build raw/, wiki/, and schema as separate layers with different write permissions.
RAG vs incremental compilation
See why query-time retrieval and ingest-time synthesis produce different maintenance behavior.
Memory lifecycle
Move claims from raw source to summary, semantic page, runbook, contradiction record, or archive.
index.md and log.md
Use two small files to make a large markdown graph navigable without a heavy vector database.