Your research disappears into chat sessions. Vaultweaver makes it compound.

Install  •   Quick Start  •   How It Works  •   Commands  •   Why Not RAG?

The Idea

Most LLM token spend goes into throwaway chat responses -- you ask, get an answer, close the tab. Nothing persists.

Vaultweaver flips this. It compiles raw sources into an interconnected wiki that grows richer with every source you add. The wiki becomes a persistent, compounding artifact -- not a disposable conversation.

What Makes It Different

Without Vaultweaver Chat answers vanish after the session 20 papers = 20 isolated summaries Re-explain context every conversation No cross-references between ideas Knowledge doesn't compound

With Vaultweaver Knowledge persists as markdown files 20 papers = interconnected knowledge graph Wiki provides full context automatically [[wiki-links]] connect everything Every source makes the whole wiki richer

Install

git clone https://github.com/Apexlify/vaultweaver.git
cd vaultweaver
bash install.sh

That's it. Start a new Claude Code session and you have /wiki.

 Manual install

cp -r . ~/.claude/skills/vaultweaver
cp SKILL.md ~/.claude/commands/wiki.md

Quick Start

/wiki init "My Research"

Drop files into raw/ -- papers, articles, notes, screenshots, anything.

/wiki ingest
/wiki compile
/wiki query "How does X relate to Y?"

Or skip the commands -- just talk:

"I dropped 3 papers in raw, process them"
"What do we know about attention scaling?"
"Check the wiki for issues"

Vaultweaver triggers automatically.

How It Works

flowchart LR
    subgraph YOU
        A[Drop files\nin raw/]
        B[Ask\nquestions]
        C[Say\n'file it']
    end

    subgraph CLAUDE
        D[Ingest\n10-15 pages\nper source]
        E[Query\nwith citations]
        F[Compound\ninto wiki]
    end

    subgraph WIKI["wiki/"]
        G[sources/]
        H[concepts/]
        I[entities/]
        J[queries/]
        K[index.md]
    end

    A --> D
    D --> G & H & I & K
    B --> E
    E --> B
    C --> F
    F --> J & K

    style YOU fill:#1a1a2e,stroke:#e94560,color:#e0e0e0
    style CLAUDE fill:#1a1a2e,stroke:#53a8b6,color:#e0e0e0
    style WIKI fill:#1a1a2e,stroke:#7c3aed,color:#e0e0e0

Loading

The Three Layers

Layer	What	Owner
raw/	Source documents -- papers, articles, notes, images	You (immutable)
wiki/	Compiled knowledge -- summaries, concepts, entities, queries	Claude (writes everything)
SKILL.md	Schema -- conventions, triggers, workflows	Vaultweaver

Commands

Command	Description	LLM
/wiki init "Topic"	Create wiki structure	--
/wiki ingest	Process raw/ sources -- 10-15 pages per source	Yes
/wiki compile	Build missing concept + entity articles	Yes
/wiki query "Q"	Answer from wiki with [[citations]]	Yes
/wiki lint	Find broken links, gaps, contradictions	Yes
/wiki search "term"	BM25 ranked search (instant, local)	--
/wiki status	File counts + recent activity	--
/wiki serve	Web search UI on localhost:5000	--

Auto-Pilot

Two hooks run automatically -- zero commands needed:

sequenceDiagram
    participant U as You
    participant C as Claude Code
    participant W as Wiki

    Note over C: Session Start
    C->>W: check-wiki-drift.sh
    W-->>C: "2 unprocessed files in raw/"
    C-->>U: Drift alert

    Note over U: During session...
    U->>W: Drop new files in raw/
    U->>C: Ask questions, do work

    Note over C: Session End
    C->>W: Auto-ingest new raw/ files
    C->>W: Update concepts, entities, index
    W-->>C: Wiki updated silently

Loading

The Ingest Pipeline

Each source goes through a 7-step pipeline, touching 10-15 wiki pages:

flowchart TD
    A[raw/paper.md] --> B[READ\nExtract claims, entities,\nconcepts, data points]
    B --> C[WRITE SOURCE\nwiki/sources/paper.md]
    C --> D[CREATE CONCEPTS\nwiki/concepts/*.md]
    C --> E[CREATE ENTITIES\nwiki/entities/*.md]
    C --> F[CROSS-REFERENCE\nUpdate existing pages]
    C --> G[UPDATE INDEX\nwiki/index.md]
    C --> H[APPEND LOG\nwiki/log.md]

    style A fill:#e94560,stroke:#e94560,color:#fff
    style B fill:#1a1a2e,stroke:#53a8b6,color:#e0e0e0
    style C fill:#1a1a2e,stroke:#7c3aed,color:#e0e0e0
    style D fill:#1a1a2e,stroke:#53a8b6,color:#e0e0e0
    style E fill:#1a1a2e,stroke:#53a8b6,color:#e0e0e0
    style F fill:#1a1a2e,stroke:#f59e0b,color:#e0e0e0
    style G fill:#1a1a2e,stroke:#10b981,color:#e0e0e0
    style H fill:#1a1a2e,stroke:#10b981,color:#e0e0e0

Loading

Why 10-15 pages? A paper mentions dozens of concepts. If you only write a summary, you get an isolated note. Create concept pages, entity pages, and cross-references, and you get a knowledge graph where every future source plugs into existing connections. Your 20th paper is 10x more valuable than your 1st.

Querying

User: /wiki query "What are the tradeoffs between
       exact and approximate attention?"

Claude: Based on [[concepts/flash-attention]],
        [[concepts/linformer]], and [[concepts/performer]]:

        | Method          | Exact? | Complexity | Memory |
        |-----------------|--------|------------|--------|
        | Full attention  | Yes    | O(n^2)     | O(n^2) |
        | Flash Attention | Yes    | O(n^2)     | O(n)   |
        | Linformer       | No     | O(n)       | O(n)   |
        | Performer       | No     | O(n)       | O(n)   |

        Want me to file this into the wiki?

User: yes

Claude: --> wiki/queries/exact-vs-approximate-attention.md

Filed answers compound. Next time you query about attention, this synthesis is already part of the wiki.

Health Checks

/wiki lint runs seven checks:

Check	What it finds	Auto-fix?
Contradictions	Conflicting claims across pages	Reports for judgment
Stale claims	Superseded by newer sources	Partial
Orphan pages	No inbound [[links]]	Adds cross-refs
Missing concepts	Mentioned often, no article	Creates stubs
Broken links	[[links]] to non-existent pages	Creates stubs
Data gaps	Partially covered topics	Suggests sources
Thin pages	Too little content	Partial

Lint tells you what to research next. It turns the wiki from a static collection into a research roadmap.

 Link normalization rules

All of these resolve to the same page:

Input	Resolves to
[[gpt-series]]	gpt-series.md
[[GPT_series]]	gpt-series.md
[[GPT Series]]	gpt-series.md
[[concepts/gpt-series]]	gpt-series.md
[[gpt-series|GPT Family]]	gpt-series.md

BM25 Search

Instant local search -- no LLM, no API calls:

python search.py wiki/ "attention mechanism"       # CLI search
python search.py wiki/ --json "transformer"        # JSON output
python search.py wiki/ --serve                     # Web UI
python search.py wiki/ --stats                     # Index stats

Web UI at localhost:5000 -- dark theme, score badges, click-through to pages.

Why Not RAG?

	RAG	Vaultweaver
Structure	Flat chunks in a vector DB	Interconnected [[wiki-links]] graph
Cross-refs	None -- chunks are isolated	Every page links to related pages
Contradictions	Hidden (chunks conflict silently)	Explicitly noted in articles
Compounding	Each query starts from scratch	Answers filed back improve future queries
Browsability	Need a special UI	Open in Obsidian or any editor
Infrastructure	Vector DB + embeddings	Plain markdown files
Best for	1K-1M documents	10-500 sources (personal research)

RAG retrieves. Vaultweaver compiles.

Wiki Structure

my-project/
|
|-- raw/                         You own this (immutable)
|   |-- paper-on-transformers.md
|   |-- blog-post-on-bert.md
|   +-- architecture-diagram.png
|
|-- wiki/                        Claude owns this
|   |-- index.md                 Categorized page catalog
|   |-- log.md                   Parseable activity log
|   |-- overview.md              High-level synthesis
|   |-- schema.md                Wiki conventions
|   |-- sources/                 One summary per raw file
|   |-- concepts/                Encyclopedic articles
|   |-- entities/                People, orgs, tools
|   +-- queries/                 Filed Q&A
|
+-- search.py                    BM25 search engine

Obsidian Integration

Open wiki/ in Obsidian for:

Graph View -- visual knowledge map showing all concepts and their [[connections]] Backlinks -- every page shows what links to it

Live Reload -- watch the wiki grow as Claude writes pages Dataview -- query YAML frontmatter across all pages

Every page has structured frontmatter:

---
title: "Self-Attention"
type: concept
tags: [attention, mechanism, transformer]
created: 2026-04-07
sources: [raw/transformers.md, raw/bert-paper.md]
---

Page Conventions

Element	Convention
Wikilinks	[[hyphenated-lowercase]] -- link liberally
First paragraph	1-2 sentence summary of the page
Citations	According to [[sources/paper-name]]...
Contradictions	Always noted explicitly, never silently resolved
Sources section	Every concept/entity page ends with ## Sources
Log entries	## [YYYY-MM-DD] operation | title -- parseable with grep

Repo Structure

vaultweaver/
|-- SKILL.md               The brain: triggers, decision tree, operations
|-- hooks.json             Auto-pilot: Stop + SessionStart hooks
|-- settings.json          Registry metadata: name, version, tags
|-- scripts/
|   |-- check-wiki-drift.sh   Detects unprocessed raw/ files
|   +-- search.py              BM25 engine + Flask web UI
|-- references/
|   +-- operations.md          Detailed 10-step workflows
|-- assets/
|   +-- banner.svg             README header graphic
|-- install.sh             One-command install
|-- uninstall.sh           Clean removal
|-- LICENSE                MIT
+-- README.md              You are here

Uninstall

bash uninstall.sh

Removes the skill. Your wiki data is never touched.

Credits

Built by Apexlify

The AI backbone powering wiki compilation

The perfect viewer for interconnected markdown

The open-source community that set the standard

---

If Vaultweaver is useful, give it a star.

Built with Claude Code. Viewed in Obsidian. Knowledge that compounds.