Knowledge engineering loop

Many knowledge bases fail not because they lack content, but because content decays: sources pile up, summaries scatter, concepts get defined twice, answers stay trapped in chat history, and a few months later it is hard to know which claims are still trustworthy.

Agent Knowledge borrows from the LLM Wiki pattern: treat a knowledge base as a continuously compiled engineering system. Humans choose sources and review important results; agents and tools handle repetitive summarization, linking, checking, and filing.

Engineering analogy

Software engineering	Agent Knowledge	Meaning
src/	sources/	Raw input and evidence from pages, meetings, papers, interviews, or internal docs.
build/	wiki/ and compiled/	wiki/ is the primary compiled artifact; compiled/ is the derived runtime view.
build logs	runs/	Compile, lint, review, eval, and health-check records.
compiler	Agent Skill, client tool, CI, or script	Reads sources and updates wiki, compiled views, and indexes.
IDE	any editor or client	This can be Obsidian, a filesystem, a web app, or a desktop client.
lint / CI	health checks and evals	Detect missing sources, contradictions, orphan pages, stale claims, and injection risks.

In personal workflows, raw/ often maps to sources/, and outputs/ can map to runs/, wiki/synthesis/, or compiled/. The standard does not require these aliases or any specific editor.

Minimal loop

A sustainable knowledge pack should run four steps:

1. Ingest sources: put raw material in sources/, preserving source URL, author, publication time, capture time, and license information.
2. Compile knowledge: incrementally compile sources into wiki/ pages such as source summaries, concepts, entities, open questions, and contradictions.
3. Derive runtime views: generate short compiled/ files from wiki/, such as facts.md, briefing.md, and boundaries.md.
4. Check and file back: write lint, health-check, eval, and useful answer artifacts back into the pack.

Turning answers into inventory

A complex answer that only remains in chat history is usually lost. Agent Knowledge recommends filing reusable answers, but not treating every output as fact.

Recommended rules:

• temporary diagnostics, compile logs, and health reports go to runs/
• reusable, sourced synthesis goes to wiki/synthesis/
• frequently loaded short conclusions can be derived into compiled/
• unconfirmed, unsourced, or disputed answers must carry status and must not enter ready runtime views

Example:

---
question: When should we use RAG instead of lightweight indexes?
asked_at: 2026-05-01
status: needs-review
sources:
  - sources/articles/local-indexing.md#L12
  - wiki/concepts/rag.md
---

# RAG and lightweight indexes

## TL;DR

For small and medium packs, prefer `wiki/index.md`, full-text search, and explicit source maps. Add vector retrieval only when scale, semantic retrieval needs, or recall requirements exceed lightweight indexes.

## Evidence

- ...

## Uncertainty

- No local benchmark yet for packs above ten thousand notes.

Health checks

Health checks are not decorative. They are the cheapest way to keep a knowledge pack trustworthy over time.

Check regularly for:

• consistency: conflicting definitions for the same concept
• completeness: important pages missing definitions, examples, or sources
• islands: pages with too few inbound or outbound links
• freshness: stale sources or claims
• traceability: important claims that cannot trace back to sources/
• security: prompt injection, secrets, or sensitive content in sources

Health-check results should be written to runs/health-<date>.json or runs/health-<date>.md. If a check finds serious issues, the maintenance tool should propose needs-review, stale, or disputed.

Do not start with RAG

Agent Knowledge does not reject RAG, but a vector database should not be the first step.

For small packs, start with:

• wiki/index.md
• wiki/concepts/
• wiki/sources/
• lightweight full-text search
• source maps

Add indexes/vector/ as a rebuildable acceleration layer only when scale, recall difficulty, or multilingual semantic search needs exceed those mechanisms. A vector index is still not fact authority.

Two-week pilot

Week one: run sources/ -> wiki/.

• Create one small knowledge pack.
• Add 5 to 10 high-quality sources.
• Compile source summaries, concept pages, and an index.
• Record the first runs/compile-...json.

Week two: run filing and checks.

• Write complex answers into wiki/synthesis/ with sources and status.
• Generate the first health-check report.
• Fix missing sources, orphan pages, and conflicting claims.
• Derive short conclusions into compiled/ only after review.

The goal is not to build a large system immediately. It is to establish a durable loop: ingest, compile, use, file back, and check.