vipin wiki

LLM Wiki

3 min read

LLM Wiki

An LLM Wiki is a knowledge-management pattern in which an LLM continuously compiles raw source material into a persistent, structured, interlinked markdown wiki.

Core Idea

The key shift is from query-time rediscovery to persistent synthesis.

Instead of retrieving raw chunks from source documents every time a question is asked, the agent incrementally updates durable pages that already contain summaries, cross-links, contradictions, and emerging synthesis.

Core Layers

  1. Raw sources: immutable source-of-truth materials.
  2. Wiki: the maintained knowledge layer.
  3. Schema: the instruction file that tells the LLM how to operate consistently.

Why It Matters

  • knowledge compounds over time
  • cross-references persist
  • contradictions can be tracked explicitly
  • useful answers do not disappear into chat history
  • maintenance work becomes cheap enough to sustain

Operational Modes

  • Ingest: read a source and integrate it into the wiki.
  • Query: answer questions from the wiki and optionally file the result.
  • Lint: inspect the wiki for gaps, stale claims, or structural issues.

Practical Extensions

  • personalization layer: summaries and prioritization should account for the reader, not only the source
  • layered context packs: load L0/L1/L2/L3 context instead of the full wiki every time
  • dual output: good work produces both a user-facing answer and durable wiki updates
  • divergence check: important pages should preserve counterarguments and data gaps
  • structured catalog and search: index files remain useful, but machine-readable cataloging helps at larger scale

LLM Wiki v2 Extensions

The pattern was extended with production lessons from the agentmemory engine (see 2026-06-19-llm-wiki-v2-pattern). The v2 modules, and how WEIPING_WIKI implements each:

  • Memory lifecycle — numeric confidence, Ebbinghaus retention decay, and explicit supersession, via the advisory wiki.py lifecycle audit over optional confidence / last_confirmed / superseded_by frontmatter.
  • Typed knowledge graph + traversal — wiki.py graph (stats/neighbors/path/export) over the catalog’s resolved links, plus typed relations parsed from the schema relation vocabulary.
  • Hybrid search — the existing BM25-lite ranking gains search --graph (1-hop neighbor fusion via reciprocal rank fusion) and search --semantic (best-effort agentmemory vector search).
  • Self-healing lint — health --fix applies non-destructive repairs (rebuild stale catalog, inject missing title/type/created); health also reports a per-page quality score.
  • Crystallization — wiki.py crystallize turns a high-value outcome into a routed, durable page with index/log/catalog updates and lint gating.
  • Privacy filter-on-ingest — wiki.py scrub flags secrets and private paths before a source becomes a page.
  • Event-driven hooks — optional scripts/hooks/session-start.py and session-end.py helpers.
  • Covered by agentmemory — consolidation tiers, mesh-sync, contradiction-heal, and graph query come from the active agentmemory layer rather than being reimplemented in the wiki.

See 2026-06-19-weiping-wiki-upgrade-audit for the full audit and verification.

Counterpoints and Gaps

  • at larger scale, handwritten markdown indexes alone may become fragile
  • graph interfaces can be helpful, but they are secondary to ingest discipline, page quality, and searchable structure
  • pure wiki compilation still benefits from selective retrieval over raw sources when the maintained layer is incomplete

Graph View

Backlinks