jmagly/knowledge-base-quickref

Knowledge Base Framework — Quick Reference

This is your always-loaded directory for the AIWG knowledge-base framework. It does not list every skill. Most heavy lifting comes from the semantic-memory kernel in aiwg-utils (memory-ingest, memory-lint, etc.) — this framework is a thin topology on top.

Canonical access pattern: discover → show

When you find a candidate via aiwg discover, fetch its body with aiwg show <type> <name>. Never use find, ls, Glob, or direct Read on <provider>/skills/ paths — those reflect the kernel-pivot deploy state, not the full surface.

aiwg discover "<phrase>"             # find — returns ranked candidates
aiwg show skill <name>               # fetch — streams the SKILL.md body

If your platform's Skill tool errors on a non-kernel skill (expected — most aren't kernel), the fallback is aiwg show, never filesystem browsing. Last-resort if aiwg itself is broken: read directly from $AIWG_ROOT/agentic/code/... (the canonical corpus, always present).

How to use this quickref

1. Identify the capability domain the user's need belongs to
2. Pick a curated phrase from that domain
3. Run aiwg discover "<phrase>" and surface the top match to the user

Do not enumerate skills from memory. Discovery is the lookup surface.

What this framework is for

A thin topology on top of AIWG's semantic-memory kernel — turning any project's .aiwg/kb/ into a queryable knowledge base. Sources get ingested into structured pages (entities, concepts, summaries, syntheses) with cross-references, deduplication, and lint coverage. Pairs naturally with the llm-wiki addon for Obsidian-compatible profiles (book-companion / personal / research-deep-dive / business-team / generic).

Capability domains

| Domain | Covers | |---|---| | KB lifecycle | Ingest sources, health-check the KB | | Semantic memory kernel (in aiwg-utils) | Generic ingest/lint/log/query primitives any consumer can declare a topology against | | LLM-wiki profiles | Topology profiles that shape how kb-ingest derives pages | | Cross-ref traversal | Graph-native via aiwg index neighbors --graph kb |

Curated discovery phrases

KB lifecycle

aiwg discover "kb-ingest"                      # → kb-ingest (score 1.00)
aiwg discover "ingest source into knowledge base" # → kb-ingest
aiwg discover "kb-health"                      # → kb-health (score 1.00)
aiwg discover "knowledge base lint"            # → kb-health

Semantic memory kernel (aiwg-utils)

aiwg discover "memory ingest"                  # → memory-ingest
aiwg discover "memory lint"                    # → memory-lint
aiwg discover "memory log append"              # → memory-log-append
aiwg discover "memory log render"              # → memory-log-render
aiwg discover "memory query capture"           # → memory-query-capture

LLM-wiki profiles (in the llm-wiki addon)

aiwg discover "llm wiki profile"               # → llm-wiki addon entries
aiwg discover "book companion knowledge base"  # → llm-wiki book-companion profile
aiwg discover "research deep dive wiki"        # → llm-wiki research-deep-dive profile

Cross-ref traversal (uses the artifact index, not a skill)

aiwg index neighbors --graph kb --node <slug>  # traverse the KB graph

How knowledge-base composes with semantic-memory

kb-ingest  ─────┐                       ┌──── memory-ingest (kernel)
                ├── declares topology ──┤
kb-health  ─────┘                       └──── memory-lint   (kernel)
                                              memory-query-capture
                                              memory-log-append / render

Every KB entry is a semantic-memory entry with a KB-specific topology (page types, cross-ref style, derived-pages config). The kernel handles ingest mechanics; this framework declares what shape the KB takes.

Page types

When ingesting via kb-ingest, the topology produces:

• Entity pages — people / orgs / products / works (one per noun)
• Concept pages — ideas / methods / principles
• Source summaries — per-source distillation (one per ingested URL/file)
• Synthesis pages — composite views across multiple sources

Cross-references between these are graph-native (visible to aiwg index neighbors).

Profile selection (via llm-wiki addon)

| Profile | Use for | |---|---| | book-companion | Reading a book, building a structured companion | | personal | Personal knowledge / journal-of-ideas | | research-deep-dive | Academic research project (uses research-corpus conventions) | | business-team | Team-shared business KB | | generic | No profile chosen — vanilla semantic-memory shape |

Install via aiwg use llm-wiki --profile <name>. The profile shapes how kb-ingest derives pages.

Artifact directory layout

.aiwg/kb/
├── entities/         # Entity pages (PROF-* compatible if research-corpus also installed)
├── concepts/         # Concept pages
├── summaries/        # Per-source distillation
├── syntheses/        # Composite views
└── log.jsonl         # Semantic-memory event log

When the curated phrases don't fit

aiwg discover "<your need, paraphrased>" --limit 5

Anti-pattern: don't enumerate

If a user asks "what KB skills are available?", do not list from this skill. Run:

aiwg discover --type skill --limit 20 "<their interest area>"