Your Claude Code sessions flow downstream. Alluvium is what settles.

Alluvium auto-archives Claude Code sessions to your Obsidian vault. After each session, it distills the transcript into knowledge — not chat logs — and merges it into your existing topic pages.

You don't save your sessions. Your sessions save themselves.

🟢 v0.1.0 — full Karpathy LLM-Wiki algorithm implemented end-to-end: wiki-aware ingest (LLM sees existing topics, reuses slugs instead of minting parallel duplicates), episode filter, alluvium lint for periodic deduplication, all backed by 327+ unit tests + 18 e2e tests. Manual install for now (Homebrew formula written but tap not yet published). See CHANGELOG.md for the full feature matrix and docs/LLM_WIKI_DOCTRINE.md for the product philosophy.

# 1. Install (pick one — the Homebrew tap is prepared but not yet published;
#    once the v0.1.0 GitHub release lands, `brew tap Tespera/alluvium` will work.
#    Until then, build from source.)
cargo install --git https://github.com/Tespera/alluvium

# or, from a local checkout:
git clone https://github.com/Tespera/alluvium
cd alluvium
cargo install --path .

# 2. Configure: vault path, recipe, API key
alluvium init

# 3. Register the Claude Code plugin (init prints this command)
claude plugin install /path/to/alluvium

# 4. (Optional) Backfill the past week of sessions
alluvium replay --since 7d

After that, just keep using Claude Code. Each session ends → distilled knowledge appears in your vault, integrated into existing topic pages (not piled up as new files). The LLM reads your wiki's existing-topics index before extracting facts, so a session about "atomic file writes" adds to the existing atomic-write page instead of creating a parallel atomic-file-write / file-atomic-write / 三向-原子写盘 zoo.

Run alluvium status to see recent archives. Run alluvium dry-run to preview the next distillation without writing. Run alluvium lint periodically — it scans the wiki for near-duplicate pages that snuck through and asks the LLM whether to merge them; pass --apply to actually do it.

Most Claude + Obsidian tools either embed Claude inside Obsidian (you ask it to do things) or expose your vault to Claude via MCP (Claude reads/writes when prompted). Alluvium does neither.

It watches Claude Code sessions end, distills the conversation, and updates a Karpathy-style LLM wiki in your vault.

Passive. No slash command, no sidebar, no "remember to save." Hook-driven; you forget it's there.

Compounding. Each session merges into existing topic pages instead of creating yet another timestamped note. Your knowledge converges instead of accumulating.

• Installed as a Claude Code Plugin (.claude-plugin/plugin.json)
• 4 hooks: SessionStart / PreCompact / Stop / SessionEnd
• Distillation via Anthropic API with externalized prompt templates (prompts/*.toml)
• Vault layout follows Karpathy's LLM Wiki pattern (raw/, wiki/, log.md, index.md)
• Single Rust binary, distributed via Homebrew

See docs/ for design details.

Alluvium calls the Anthropic API once per Claude Code session. Default model is Claude Haiku 4.5 (cheapest tier).

Rough estimate (default dev-journal recipe, ~50-message session):

Switch to Sonnet (better notes, ~5× cost) or another model in ~/.config/alluvium/config.toml. The verbose recipe roughly doubles cost; minimalist halves it.

See docs/PRIOR_ART.md. We learned specifically from obsidian-second-brain, claude-obsidian, and cognee-integrations.

Dual-licensed under the MIT License and Apache License 2.0, at your option. This matches Rust ecosystem conventions.