<aside> 📌 Every time you open Claude Code, it starts from scratch. It re-reads your files, re-discovers your project, and burns through tokens figuring out what it already knew yesterday. Here's how to fix that permanently.

</aside>

Why Your Current Setup Is Costing You

Here's what happens in a normal Claude Code session right now:

• Claude loads your CLAUDE.md (hundreds of tokens)
• You ask a question about your project
• Claude greps through your files, reads 5-10 of them, tries to piece together an answer
• That's thousands of tokens burned on discovery — not on actual work

Now multiply that by every conversation you have. You're paying for Claude to re-learn your project from scratch every single time. The fix: make Claude build and maintain a wiki that it reads from instead. One index scan, one targeted page read, done. That's 60-70% fewer tokens on knowledge lookups.

What You're Building

An Obsidian vault that Claude Code maintains for you:

• A wiki of interlinked markdown pages — one page per topic, project, concept, or person. Claude writes and updates all of them.
• A "hot context" file (hot.md) — a rolling snapshot of the last few days. What you're working on, what changed, what Claude needs to know right now.
• An index file — a lightweight table of contents Claude scans to find exactly what it needs.
• Tiered retrieval — wiki summaries first (fast, cheap), raw source files when it needs the full picture.
• Conversation logging — every session gets a timestamped summary. Claude remembers last Tuesday.
• Automatic archiving — stale pages get flagged for your approval. Wiki stays lean.
• Obsidian's graph view — visual map of everything Claude knows.

The Real Unlock: Tiered Retrieval

The wiki doesn't replace your original files — it sits in front of them. Think of it like a cache:

Tier 1 — hot.md (every session, ~500 tokens). What you're working on. Last few days. Major changes. Read first, every time.