A knowledge vault is a folder of plain Markdown notes with typed frontmatter, explicit links, and a machine-checkable schema. It compounds when it is atomic, sourced, and audited, and rots when it becomes long documents nobody links.
Architecture
- vault-architecture: Numbered folder tree, scoped note types,
vault-spec.yamlas source of truth, generation over hand-building. - atomic-notes: One idea per note, claim titles, 100 to 500 words, three to seven explained links, deliberate promotion.
- vault-frontmatter-schema: Per-type required fields, empty arrays over missing keys, framework extensions.
- linking-and-tags: Links carry meaning, tags carry state, block references, MOCs grown from clusters.
- rigor-frameworks: Map the domain to a framework (MECE, ADR, IRAC, and more), layer, or compose one.
Evidence and decisions
- source-verification: One source note per work, dated verification, the corroboration gate.
- decision-journals: Journal at decision time, freeze the pre-decision section, pre-mortem, review by class.
Operations
- vault-orchestration: The capture-extract-link-audit-promote loop, the fixed agent prompts, bulk-write guardrails.
- vault-audit: The deterministic auditor, rule catalog (C, S, M series), triage, autofix tier, re-run delta.
- semantic-audit: The fixed-rubric LLM judge for defects code cannot detect.
- vault-maintenance: Weekly, monthly, and quarterly reviews, Dataview gates, the maintenance budget.
- vault-evolution: Add, deprecate, migrate; never rebuild. Includes the migration for a newly required field.
- team-vaults: Git as substrate, conflict-free conventions, one naming authority, owned decisions.
- vault-pull-request-review: Auditor as merge check, severity mapping, squash rules for atoms vs decisions.
How-to
- bootstrap-a-research-vault: Generate a vault from a spec, run the auditor, close the autofix loop.