A general-purpose, LLM-maintained personal knowledge base, built on the pattern from Karpathy's LLM-wiki gist.
Instead of re-retrieving raw documents at query time, an LLM agent incrementally builds and maintains a persistent wiki. Knowledge compounds across sessions: you supply sources and good questions; the agent does the summarizing, cross-referencing, and consistency bookkeeping.
The wiki is domain-general — it can hold knowledge about anything (people, ideas, organizations, events, …). Deep, specialized subjects are added as domain forks (e.g. AI) that layer extra rules and page types on top of the general core, without polluting it.
Storage is plain markdown + git — readable in any editor, no special tooling required.
- Raw layer (
raw/) — captured source material, in two stages:raw/inbox/— originals in native form (yyyy-mm-dd-name.{html,pdf,…}). Immutable, and git-ignored (local staging only).raw/sources/— a faithful markdown copyedit, 100% true to the original. The canonical, version-controlled record. The only edits allowed: Canadian-English spelling, meaning-preserving rephrasing, grammar fixes, and repairing broken lists/tables/formulas.
- Wiki layer (
wiki/) — extracted, cross-referenced knowledge in typed folders. All interpretation and synthesis lives here. - Schema layer —
CLAUDE.mddefines the conventions and workflows.AGENTS.mdpoints non-Claude agents to it.
CLAUDE.md schema + workflows (the agent's instructions)
AGENTS.md cross-agent pointer to CLAUDE.md
index.md catalog of all wiki pages, grouped by type
log.md append-only activity log (ingests, queries, lints)
raw/
inbox/ (git-ignored) originals + assets
sources/ faithful markdown copyedits + renamed assets
wiki/ general-purpose core types
people/ orgs/ places/ events/ concepts/
entities/ topics/ queries/ synthesis/
domains/
ai/ example domain fork (rules + material only; pages go to root wiki/)
The agent operates through three commands, defined in CLAUDE.md:
- Ingest —
add <url-or-file>: capture the original → produce a faithful copyedit → extract knowledge intowiki/, cross-linking pages and updatingindex.mdandlog.md. - Query —
ask <question>: search the wiki and synthesize a cited answer, optionally filing it back intowiki/queries/. - Lint —
lint: scan for contradictions, stale claims, orphaned pages, broken cross-references, and knowledge gaps.
A domain fork (domains/<name>/) bolts on a deep sub-domain without polluting
the general schema. It holds rules and reference material only — a
SCHEMA.md the agent layers on top of CLAUDE.md during ingest. It has no
wiki pages of its own; every page it produces goes to the root wiki/, with
its extra types becoming additional wiki/ subfolders. See
domains/ai/SCHEMA.md for a worked example.
- Slugs: lowercase-kebab-case filenames.
- Dates: ISO
yyyy-mm-dd. - Spelling: Canadian English.
- Faithfulness:
raw/sources/stays 100% true to the original; synthesis lives only inwiki/. - Cross-references: bidirectional relative markdown links.
- Point your agent at this repo (it reads
CLAUDE.mdautomatically). add <url>to ingest your first source.ask <question>to query what you've built.
The full design rationale lives in
docs/superpowers/specs/.
MIT.