I connected llm-wiki-compiler as the knowledge layer beneath my agent stack and it finally stopped repeating the same research every session.

Pattern: ingest docs → compile into interlinked markdown wiki → agent queries via MCP. Tags carry over. Map of content auto-generates.

When the agent answers something new, query --save pushes it back into the wiki as a page. Next query is smarter because the artifact is richer.

That's the difference between a stateless file upload and a knowledge base that actually accumulates.

If you're building with Hermes / Claude Code / Codex and hitting the "it doesn't know my domain" problem, a persistent markdown wiki underneath changes everything.