agentprivacy-wiki-llm-knowledge-base
The LLM-Wiki pattern: instead of re-deriving knowledge from raw documents on every query (RAG),
an LLM incrementally builds and maintains a persistent, interlinked wiki that sits between you and
the sources. Knowledge is compiled once and kept current — the cross-references are already there,
the contradictions already flagged, the synthesis already reflects everything read. The wiki is a
compounding artifact. You curate sources, explore, and ask questions; the LLM does the summarizing,
cross-referencing, filing, and bookkeeping. Obsidian (or a FedWiki) is the IDE; the LLM is the
programmer; the wiki is the codebase.
The three layers
| Layer |
What it is |
Rule |
| Raw sources |
the curated source documents (articles, papers, repos, data) |
immutable — the LLM reads, never modifies; the source of truth |
| The wiki |
the LLM-generated markdown: summaries, entity/concept pages, comparisons, an overview |
the LLM owns it entirely — creates, updates, cross-references, keeps consistent; you read it |
| The schema |
the config doc (CLAUDE.md / AGENTS.md) telling the LLM the structure, conventions, workflows |
the key file — makes the LLM a disciplined maintainer, not a generic chatbot; co-evolved over time |
The three operations
- Ingest — drop a source in; the LLM reads it, discusses takeaways, writes a summary page, updates the
index, updates the entity/concept pages it touches (a single source may touch 10–15 pages), appends a log
entry. One source at a time with supervision, or batch with less.
- Query — ask against the wiki; the LLM finds relevant pages, reads, synthesizes a cited answer. Good
answers are filed back into the wiki as new pages — explorations compound just like ingested sources.
- Lint — periodically health-check: contradictions between pages, stale claims newer sources supersede,
orphan pages (no inbound links), important concepts lacking a page, missing cross-references, data gaps to
fill with a search. Keeps the wiki healthy as it grows.
Two navigation files
- index.md — content-oriented: a catalog of every page (link + one-line summary + optional metadata),
grouped by category. The LLM reads it first when answering. Avoids embedding-RAG at moderate scale (~100s of pages).
- log.md — chronological: append-only record of ingests/queries/lints. Use a consistent prefix
(
## [2026-04-02] ingest | Title) so it's greppable (grep "^## \[" log.md | tail -5).
The load-bearing rules
- The LLM owns the wiki layer; the human owns sourcing, direction, and questions.
- Raw sources are immutable — never edited, always the source of truth.
- The schema governs — disciplined maintenance comes from the config doc, not the model's whim.
- Maintenance cost is near zero (the LLM touches 15 files without boredom), so the wiki stays current —
that is why it works where human-maintained wikis are abandoned.
- Answers worth keeping are filed back, not lost to chat history.
This federation IS an instance of the pattern
The agentprivacy guide federation (see the agentprivacy-wiki-sync skill) is a worked LLM-Wiki:
| LLM-Wiki concept |
agentprivacy federation realization |
| Raw sources (immutable) |
the git canon repos — agentprivacy-skills-v5, cityofmages, agentprivacy-docs, spellweb (gate G0: git is upstream, never authored-first on the wiki) |
| The wiki (LLM-owned) |
the FedWiki federation — skill / tomes / research / atlas ≈ 1,175 forkable, cross-linked pages |
| The schema (CLAUDE.md) |
the agentprivacy-wiki-sync skill — conventions, the gates G0–G7, the manifest-first discipline |
| Ingest |
the build-* sync passes — manifest-first G1 classify, project canon → pages + assets, cross-link |
| Query |
FedWiki search + activity + the federationmap; answers filed back via file-answer.js → notes/ → the The Notes hub on the guide (so explorations compound) |
| Lint |
the leak-scan + link-integrity sweep + the coverage audit + forkback.js (G7) |
| index.md |
the hub/welcome-visitors pages + the live federationmap (sized by page count) |
| log.md |
log.js → a single greppable Federation Log (grep "^## \[" log.md), published as a page; the dated chronicles remain the long-form prose log |
| cross-references |
[[wikilinks]] derived from the spellweb graph edges (G5); cross-site reference items |
| search/CLI tool (qmd) |
the FedWiki search plugin + the verify recipes (sitemap / link-integrity / leak-scan) |
The difference our framing adds: the distribution gates make the LLM-Wiki governed — what crosses
from raw→wiki (G0–G6) and wiki→raw (G7) is gated for classification, coherence, attribution, and a
First-Person sign-off, rather than ungoverned LLM edits. The Memex maintenance problem Vannevar Bush
couldn't solve — who does the bookkeeping — the LLM solves; the gates keep it honest.
Fully realised (2026-06-27): every operation now has a home — including the two the essay names that a
chronicle-only system lacked: answers filed back (file-answer.js → the The Notes hub) and a single
greppable log.md (log.js → the Federation Log page). Worked synthesis: Governing an LLM-Wiki: the
eight gates (filed in The Notes). Tooling: ~/.wiki/skill-fedwiki/{log.js, file-answer.js, build-guide-extras.js}.
When applying the pattern fresh
- Define the three layers for the domain (where raw sources live, where the wiki lives, the schema doc).
- Write the schema first (it is the LLM-Wiki's
CLAUDE.md) — conventions, page formats, the ingest/lint
workflows. Co-evolve it.
- Ingest one source: summary page → index → touched entity/concept pages → log entry.
- Query, and file good answers back.
- Lint regularly (contradictions / stale / orphans / missing pages / gaps).
- Keep
index.md (content) and log.md (chronological, greppable) current.
Upstream: the LLM-Wiki pattern essay supplied by the user (skill.localhost/llm-wiki.md) — vendored +
re-framed; the pattern is the source's, the federation mapping and gate correspondence are ours.
Verify: agentprivacy.ai · the agentprivacy-wiki-sync skill (the instantiation) ·
plan ~/.claude/plans/moonlit-stirring-lagoon.md
1---2name: agentprivacy-wiki-llm-knowledge-base3description: Building and maintaining a persistent, LLM-maintained knowledge base — the LLM-Wiki pattern: an LLM incrementally compiles immutable raw sources into a cross-linked, self-consistent wiki it owns, governed by a schema, via ingest / query / lint, with a content index and a chronological log. Activates when designing a knowledge base, ingesting sources into a wiki, health-checking one, or mapping the pattern onto the agentprivacy federation (which IS an instance of it). Part of the git-less onboarding/distribution layer kept by the Librarian 🗃️ in the Tower's Wikis.4license: Apache-2.05---67# agentprivacy-wiki-llm-knowledge-base89The **LLM-Wiki pattern**: instead of re-deriving knowledge from raw documents on every query (RAG),10an LLM **incrementally builds and maintains a persistent, interlinked wiki** that sits between you and11the sources. Knowledge is **compiled once and kept current** — the cross-references are already there,12the contradictions already flagged, the synthesis already reflects everything read. The wiki is a13*compounding artifact*. You curate sources, explore, and ask questions; the LLM does the summarizing,14cross-referencing, filing, and bookkeeping. *Obsidian (or a FedWiki) is the IDE; the LLM is the15programmer; the wiki is the codebase.*1617## The three layers1819| Layer | What it is | Rule |20|---|---|---|21| **Raw sources** | the curated source documents (articles, papers, repos, data) | **immutable** — the LLM reads, never modifies; the source of truth |22| **The wiki** | the LLM-generated markdown: summaries, entity/concept pages, comparisons, an overview | the **LLM owns it entirely** — creates, updates, cross-references, keeps consistent; you read it |23| **The schema** | the config doc (CLAUDE.md / AGENTS.md) telling the LLM the structure, conventions, workflows | the key file — makes the LLM a *disciplined maintainer*, not a generic chatbot; co-evolved over time |2425## The three operations2627- **Ingest** — drop a source in; the LLM reads it, discusses takeaways, writes a summary page, updates the28 index, updates the entity/concept pages it touches (a single source may touch 10–15 pages), appends a log29 entry. One source at a time with supervision, or batch with less.30- **Query** — ask against the wiki; the LLM finds relevant pages, reads, synthesizes a cited answer. **Good31 answers are filed back into the wiki as new pages** — explorations compound just like ingested sources.32- **Lint** — periodically health-check: contradictions between pages, stale claims newer sources supersede,33 orphan pages (no inbound links), important concepts lacking a page, missing cross-references, data gaps to34 fill with a search. Keeps the wiki healthy as it grows.3536## Two navigation files3738- **index.md** — *content-oriented*: a catalog of every page (link + one-line summary + optional metadata),39 grouped by category. The LLM reads it first when answering. Avoids embedding-RAG at moderate scale (~100s of pages).40- **log.md** — *chronological*: append-only record of ingests/queries/lints. Use a consistent prefix41 (`## [2026-04-02] ingest | Title`) so it's greppable (`grep "^## \[" log.md | tail -5`).4243## The load-bearing rules44451. The LLM owns the wiki layer; the human owns sourcing, direction, and questions.462. Raw sources are immutable — never edited, always the source of truth.473. The schema governs — disciplined maintenance comes from the config doc, not the model's whim.484. Maintenance cost is near zero (the LLM touches 15 files without boredom), so the wiki stays current —49 that is *why it works* where human-maintained wikis are abandoned.505. Answers worth keeping are filed back, not lost to chat history.5152## This federation IS an instance of the pattern5354The agentprivacy guide federation (see the `agentprivacy-wiki-sync` skill) is a worked LLM-Wiki:5556| LLM-Wiki concept | agentprivacy federation realization |57|---|---|58| **Raw sources** (immutable) | the git canon repos — `agentprivacy-skills-v5`, `cityofmages`, `agentprivacy-docs`, `spellweb` (gate **G0**: git is upstream, never authored-first on the wiki) |59| **The wiki** (LLM-owned) | the FedWiki federation — `skill / tomes / research / atlas` ≈ 1,175 forkable, cross-linked pages |60| **The schema** (CLAUDE.md) | the `agentprivacy-wiki-sync` skill — conventions, the gates G0–G7, the manifest-first discipline |61| **Ingest** | the `build-*` sync passes — manifest-first **G1** classify, project canon → pages + assets, cross-link |62| **Query** | FedWiki `search` + `activity` + the `federationmap`; **answers filed back** via `file-answer.js` → `notes/` → the **The Notes** hub on the guide (so explorations compound) |63| **Lint** | the leak-scan + link-integrity sweep + the coverage audit + `forkback.js` (**G7**) |64| **index.md** | the hub/`welcome-visitors` pages + the live `federationmap` (sized by page count) |65| **log.md** | `log.js` → a single greppable **Federation Log** (`grep "^## \[" log.md`), published as a page; the dated chronicles remain the long-form prose log |66| **cross-references** | `[[wikilinks]]` derived from the spellweb graph edges (**G5**); cross-site `reference` items |67| **search/CLI tool (qmd)** | the FedWiki `search` plugin + the verify recipes (sitemap / link-integrity / leak-scan) |6869The difference our framing adds: the **distribution gates** make the LLM-Wiki *governed* — what crosses70from raw→wiki (G0–G6) and wiki→raw (G7) is gated for classification, coherence, attribution, and a71First-Person sign-off, rather than ungoverned LLM edits. The Memex maintenance problem Vannevar Bush72couldn't solve — *who does the bookkeeping* — the LLM solves; the gates keep it honest.7374**Fully realised (2026-06-27):** every operation now has a home — including the two the essay names that a75chronicle-only system lacked: **answers filed back** (`file-answer.js` → the *The Notes* hub) and a single76**greppable `log.md`** (`log.js` → the *Federation Log* page). Worked synthesis: *Governing an LLM-Wiki: the77eight gates* (filed in The Notes). Tooling: `~/.wiki/skill-fedwiki/{log.js, file-answer.js, build-guide-extras.js}`.7879## When applying the pattern fresh80811. Define the **three layers** for the domain (where raw sources live, where the wiki lives, the schema doc).822. Write the **schema** first (it is the LLM-Wiki's `CLAUDE.md`) — conventions, page formats, the ingest/lint83 workflows. Co-evolve it.843. **Ingest** one source: summary page → index → touched entity/concept pages → log entry.854. **Query**, and file good answers back.865. **Lint** regularly (contradictions / stale / orphans / missing pages / gaps).876. Keep `index.md` (content) and `log.md` (chronological, greppable) current.8889**Upstream:** the LLM-Wiki pattern essay supplied by the user (`skill.localhost/llm-wiki.md`) — vendored +90re-framed; the pattern is the source's, the federation mapping and gate correspondence are ours.9192**Verify:** [agentprivacy.ai](https://agentprivacy.ai) · the `agentprivacy-wiki-sync` skill (the instantiation) ·93plan `~/.claude/plans/moonlit-stirring-lagoon.md`
mitchuski