Why I Built Yet Another AI Knowledge Tool (And How It's Different)

There's no shortage of knowledge management tools for AI coding agents. Graphify builds knowledge graphs. Claude-mem replays sessions. LLM Wiki compiles research wikis. Ars Contexta generates personal thinking systems from interviews.

I tried them. They're good tools. But none of them solved my specific problem.

The Problem

I use Claude Code daily on production apps. The pain isn't "Claude doesn't understand my codebase" — Graphify handles that well. The pain is knowledge decay between sessions.

That authentication bug I debugged on Monday? Forgotten by Tuesday. The decision to use JWT over sessions? Gone. The architecture of the payment flow? Rediscovered from scratch.

Session replay tools like Claude-mem capture everything, but try finding that one critical insight in a sea of compressed session logs. It's built for the AI to search, not for you to browse.

Knowledge graph tools like Graphify are snapshots — brilliant one-time analysis, but they don't grow when you ship a fix at 2am.

Wiki tools like LLM Wiki are source-driven — great for research, disconnected from your commit history.

What I Actually Needed

A system that:

1. Grows from my git workflow — not from sources I manually feed it
2. I can browse myself — not just the AI
3. Stays current automatically — catches decay before it misleads
4. Has zero dependencies — no Python, no SQLite, no vector DB, no MCP servers

Spine

Spine is a Claude Code plugin that bridges AI memory to an Obsidian vault. Knowledge is organized by what actually matters in engineering: which repo, which feature, what type of work.

Your vault/
  └─ your-repo/
       └─ Authentication/
            ├─ Authentication.md              ← spine note (the hub)
            ├─ 2026-03-18 Fix - Cookie Expiry.md
            ├─ Architecture - OAuth Flow.md
            └─ Decision - JWT vs Session.md

Navigation is three hops:

Claude Memory          →  Spine Note           →  Specific Doc
(one-line signpost)       (feature hub)           (full context)

No embeddings. No vector search. Just structured markdown with wikilinks and frontmatter tags.

How It Compares

The Self-Developing Part

This isn't a tool you have to feed. Spine grows from your work.

After you commit, /spine-capture reads your changes and auto-drafts an Obsidian doc — root cause, code snippets, wikilinks to related docs. You review and approve.

/spine-health audits the vault for undocumented work, stale docs, broken links, and tag issues. The post-commit hook nudges you when something significant lands.

Every doc gets a type tag (type/fix, type/feature, type/architecture, type/plan, type/decision) that drives Obsidian's color-coded graph view. Open the graph in Obsidian and you can see your knowledge — red nodes for bugs, green for features, purple for architecture decisions.

Why This Matters

The vault compounds. Every doc makes the next session faster. Every wikilink makes discovery easier. Every spine note means Claude starts with context instead of starting from scratch.

This is compound interest for engineering knowledge. And the cost of maintenance is near zero — because the AI handles the bookkeeping.

Your job: curate and think.
The AI's job: everything else.

Try It

claude plugin marketplace add Nodewarrior/spine
claude plugin install spine

Then in Claude Code:

/spine-init ~/Documents/MyVault

Open the vault in Obsidian, press Cmd+G, and watch the graph light up.

MIT licensed. Give it a spin — I'm actively building this and want to hear what you'd change.

GitHub → Nodewarrior/spine

If you're using Spine or have ideas for what it should do next, I'd love to hear about it in the comments.