# Research Baseline

The implementation baseline is:

- Kochim LLM Wiki implementation guide: `https://kochim.com/boards/posts/1779116400000_llm-wiki-implementation-guide-2026/`
- the project design research notes

Important conclusions:

- The durable pattern is `raw sources -> wiki -> schema`.
- `raw/` and `wiki/` must not be mixed.
- `index.md` should be a small navigation surface.
- `memory.md` should be an even smaller hot index for the facts most likely to matter in the next turn.
- `log.md` should be append-only.
- Good answers should be written back into the wiki instead of remaining in chat history.
- Duplicate and stale pages are major failure modes.
- Full automation without human review is risky for large imports or high-stakes decisions.
- Obsidian, Graphify, MCP, and RAG can be useful extensions, but v1 must work with plain Markdown and hooks.
- Retrieval should combine lexical search with explicit Markdown links before adding heavier external infrastructure.
- Consolidation should be conservative: update generated blocks and navigation surfaces, but preserve human/agent-written page bodies.
- Separate LLM API integration is not required for the current design. Claude Code and Codex already provide the reasoning/writing layer; the kit should supply context, durable files, guardrails, and safe automatic project-rule updates.
- Heavy layers such as embeddings, graph databases, MCP tools, and hierarchical summary systems are future considerations only when project wikis become too large for the plain Markdown + hook workflow.