LLM-Wiki-v3

A general-purpose memory substrate that compounds knowledge over time. Successor to the LLM-wiki pattern, with the discipline needed to keep a wiki healthy as it scales. It serves two readers from one store:

• Humans — browsing in Obsidian, a CLI, or any markdown viewer.
• LLM agents — retrieving context via an MCP server.

The wiki is domain-agnostic. You instantiate it for your domain by writing a schema (see schema/); the system makes no domain assumptions.

Principles (the whole point)

• Markdown + git is the source of truth. Every fact lives in a markdown

file under version control. Vectors, BM25, and the entity graph are derived and rebuildable from markdown via rebuild. If it can't be rebuilt from markdown, it doesn't exist.

• No numeric confidence scores. Trust is reconstructed on demand from a

provenance chain — source_id, span, extracted_at, extractor — never stored as a float.

• No retention decay. Old facts don't fade. They are superseded

(superseded_by in frontmatter); old pages stay, queries prefer current, and git records what was true when.

• Evals gate changes. No retrieval change ships without an NDCG/MRR delta;

no ingest change without a faithfulness delta (Phase 2).

• Two write paths. Curated (human-reviewed, default) and autonomous (agent

→ pending/ → eval gate → knowledge/).

• Schema is a versioned artifact; provenance is non-negotiable;