claude-wiki

Installable Python package providing Claude Code hooks and a CLI for a personal knowledge base. It captures context from your AI coding sessions automatically, compiles it into a wiki of atomic, cross-linked articles, and feeds the relevant index back into your next session — so knowledge compounds instead of evaporating.

Adapted from Karpathy's LLM Knowledge Base architecture.

Features

• Automatic capture. Session hooks extract the meaningful turns of each Claude Code session into immutable daily logs — no manual note-taking.
• Compile to atomic articles. Daily logs are distilled by an LLM into one article per concept, plus cross-cutting connections/ and filed qa/ answers, each with YAML frontmatter and provenance backlinks. --max-logs caps the cost of a run.
• Index-guided query (no RAG). query reads the catalog first, then answers from the KB — retrieval grounded in structure rather than embeddings.
• Health checks. lint runs structural checks (broken wikilinks, orphan pages, sparse articles, frontmatter schema, and catalog↔article completeness) and optional LLM contradiction checks, with --fix for safe auto-repairs and machine-readable --json output and a stable exit-code contract for CI.
• Topology & health monitoring. graph reports the link topology — orphans, hubs, and connected components — so you can spot a fragmented KB at a glance; status diagnoses repository health with --json for CI gating.
• Obsidian-friendly output. Wikilinks, frontmatter, and a per-repo {repo_name}.md catalog keep the vault clean for Graph view; a global core.md registry links every registered repo with [[owner/repo]] links.
• Project or user mode. Keep the KB inside the repo (.claude/knowledge/) or in an XDG user-wide vault; migrate moves data between modes with rollback.
• Fast, non-fatal hooks. Handlers do only local I/O within the Claude Code timeout and offload LLM work to a backgrounded flush process; failures are logged, never fatal.

How it works

Claude Code session event
  -> claude-wiki-hook <Event>      (SessionStart / SessionEnd / PreCompact)
  -> hook handler (fast local I/O) -> backgrounded flush -> daily/YYYY-MM-DD.md

claude-wiki compile
  -> daily logs -> LLM -> kb_root/concepts/ · connections/ · qa/ + catalog

SessionStart hook -> injects catalog + recent daily log into next session

Daily logs are the append-only source of truth; the compiled wiki is LLM-owned and rebuilt from them. The catalog is the primary retrieval mechanism — the SessionStart hook injects it into each new session so context carries forward.

Install

With uv (recommended):

uvx claude-wiki init

From source in a clone of this repo:

uv sync --frozen

Requires Python 3.12+. No API key — LLM calls use Claude Code's own credentials.

Usage

Initialize a repository:

claude-wiki init

Daily commands:

claude-wiki compile [--all] [--file FILE] [--dry-run] [--continue-on-error] [--max-logs N] [--path PATH]
claude-wiki query "your question" [--file-back] [--json] [--path PATH]
claude-wiki lint [--structural-only] [--fail-on-warning] [--fix] [--threshold N] [--json] [--path PATH]
claude-wiki graph [--json] [--top N] [--path PATH]
claude-wiki status [--json] [--path PATH]
claude-wiki tags [--json] [--path PATH]
claude-wiki migrate [--dry-run] [--path PATH] [--kb-dir KB_DIR] [--daily-dir DAILY_DIR]
claude-wiki rename-catalog [--dry-run] [--path PATH]

query, lint, status, and graph accept --json for machine-readable output and follow a stable exit-code contract. See the CLI reference for every flag.

Hook entry points (called by Claude Code via .claude/settings.local.json by default):

claude-wiki-hook SessionStart
claude-wiki-hook SessionEnd
claude-wiki-hook PreCompact

What claude-wiki init creates

my-project/
├── .claude-wiki.lock              # per-repo config (machine-managed)
├── .claude/settings.local.json   # repo-local hook registration (default)
└── .claude/daily/                 # conversation logs (created on first flush)

Use claude-wiki init --global to write hooks to ~/.claude/settings.json instead. Use claude-wiki init --path PATH to target a different repository root, or --no-hooks to skip hook installation.

Configuration

.claude-wiki.lock fields:

{
  "repo_name": "my-project",
  "repo_owner": "local",
  "layout_version": "2",
  "kb_dir": "project",
  "daily_dir": ".claude/daily",
  "reports_dir": "reports",
  "timezone": "UTC",
  "compile_after_hour": 18
}

• layout_version tracks the internal directory-layout generation. New repositories use "2".
• kb_dir is project (repo-relative .claude/knowledge/), user (XDG vault ~/.local/share/claude-wiki-vault/<owner>/<repo>/), or a custom path.
• daily_dir defaults to .claude/daily in project mode and ~/.local/share/claude-wiki-daily/<owner>/<repo>/ in user mode.
• reports_dir is deprecated; reports are written to the cache directory (<repo>/.claude/reports/ in project mode).

Environment overrides: CLAUDE_WIKI_PROJECT_DIR (KB location), CLAUDE_WIKI_STATE_DIR (machine state), CLAUDE_WIKI_CACHE_DIR (reports), and CLAUDE_WIKI_DEBUG (verbose hook logging).

Documentation

• Full docs in docs/ — tutorials, how-to guides, and reference.
• CHANGELOG.md for release history.
• docs/adr/ for architecture decisions.

Community

• Contributing — branch naming, PR checklist, dev setup.
• Code of Conduct — Contributor Covenant 2.1.
• Security policy — responsible disclosure.

Development

make dev              # install with dev dependencies
make install-precommit # install git hooks (run once per clone)
make test             # run pytest
make lint             # ruff check
make format           # ruff format + mdformat
make typecheck        # mypy
make precommit        # all pre-commit hooks
make all              # full CI gate (format, lint, typecheck, test, precommit)