Mneme wiki

Mneme wiki — the knowledge tier

The cache (Tier 1) is lean and injected into every chat. A wiki (Tier 2) is the opposite: a per-corpus knowledge base built from your documents, read on demand and never injected. So a wiki can be large — 10–15 pages per source is fine. This is the gist's "LLM Wiki": maintained markdown over raw documents.

Where a wiki lives

• Home: dirname(<global cache>)/wiki/ (a sibling of cache/), override MNEME_WIKI_DIR. Project overlay: <project>/.mneme/wiki/.
• One corpus per subdir: wiki/<corpus>/
  • index.md — the corpus table of contents (Pages / Entities / Concepts). Hand-maintained.
  • pages/<slug>.md — summary, entity, and concept pages. Frontmatter like a cache note, but bodies may be long. Cross-link with [[slug]].
  • sources/ — the raw (and markitdown-converted) source files.
  • log.md — append-only timeline (ingest, recall-filed, lint-fix), via lib/log.sh. Never injected.

The loop

1. Ingest — /mneme:ingest <source> [--wiki <name>] brings a document in, summarizes it into pages, and updates index.md.
2. Query — /mneme:recall "<q>" --wiki <name> reads index.md first, then the relevant pages, and answers with citations. Plain /recall never touches a wiki.
3. Audit — /mneme:lint --wiki <name> runs the link checks (dead [[links]], orphans) over the corpus's pages.

Shared engine

A wiki reuses lib/ (the same as the cache): log.sh for log.md, links.sh for [[link]] graphs between pages, md.sh for frontmatter, wiki.sh for home/corpus resolution. Run helpers under explicit bash (the agent's shell may be zsh). The cache and the wiki are two tiers on one engine.

The invariant

A wiki is never injected. The SessionStart loader emits at most one line naming the corpora (Wikis: <names> — query with /recall --wiki <name>); it never reads a wiki's index.md or pages. Knowledge enters a chat only when you query it.