Claude Code + Obsidian + Graphify: The Definitive Guide to Token Savings & Persistent Memory

Persistent memory across sessions with Obsidian Zettelkasten and selective code context with Graphify. Token savings depend on the task; the historical 71.5x/session and 499x/query claims are author-reported and not reproducible from published data.

A complete setup to turn Claude Code into an agent with long-term memory and full codebase awareness — without wasting tokens re-reading files.

🇧🇷 Leia em Português

Table of Contents

1. The Problem
2. The Solution (Overview)
3. Part 1 — Obsidian as Persistent Memory
4. Part 2 — Chat Import Pipeline
5. Part 3 — Graphify (Codebase Knowledge Graph)
6. Part 4 — Complete Workflow
7. Part 5 — Auto-Save Hook (Safety Net)
8. Real Results
9. Troubleshooting

The Problem

When working with Claude Code, two problems silently eat your tokens:

Problem 1 — Amnesia between sessions. Every time you open a new session, you have to re-explain your project: stack, past decisions, current bugs, what's left to do. Claude Code remembers nothing from the previous session.

Problem 2 — Codebase re-reading. Repeated source exploration can consume context tokens, but how much depends on the task and navigation tools. The historical illustration compares ~20,000 tokens of source context with ~280 tokens of graph query output: approximately 71.4x less retrieval context, not a measured reduction in complete session usage. It assumes the smaller output answers the same question; the original payloads and tokenizer were not published.

The Solution (Overview)

Two complementary systems, each solving a different problem:

Obsidian handles what was decided (declarative memory). Graphify handles how the code is structured (structural map). Together, they can help Claude Code recover relevant context. Source reads remain necessary when the graph or notes are incomplete or stale.

Part 1 — Obsidian as Persistent Memory

Concept

A single, centralized Obsidian vault acts as Claude Code's "second brain." It stores decisions, context, progress, and knowledge for all your projects. Notes follow the Zettelkasten method: atomic (one idea per note), densely interlinked, with standardized metadata.

Claude Code accesses the vault through CLAUDE.md and custom skills.

Recommended Structure

~/vault/                              # SINGLE vault for all projects
├── CLAUDE.md                         # global instructions for Claude Code
├── permanent/                        # consolidated atomic notes
├── inbox/                            # raw capture (ideas, drafts)
├── fleeting/                         # quick temporary notes
├── templates/                        # note templates
├── logs/                             # global session logs
├── references/                       # reference material
├── my-project/                       # MOCs and notes for project X
│   ├── architecture/                 #   architecture, decisions, conventions
│   ├── pipeline/                     #   data flows, APIs
│   ├── data/                         #   schema, data model
│   ├── features/                     #   planned/implemented features
│   └── logs/                         #   project session logs
├── another-project/                  # MOCs and notes for project Y
│   └── ...
├── chats/                            # imported Claude chats
│   ├── code/                         #   from Claude Code
│   └── web/                          #   from Claude Web/App
└── graphify/                         # codebase knowledge graphs
    ├── my-project/                   #   graph notes for project X
    └── another-project/              #   graph notes for project Y

Why a single vault? Having one vault per project fragments knowledge. With a single vault, a note about "Supabase Auth" links to both project A and B. The graph view reveals cross-project connections you didn't expect.

Step-by-Step Setup

Prerequisites:

• Claude Code installed and authenticated
• Obsidian installed (free: obsidian.md)

1. Create the vault:

Obsidian → "Create new vault" → choose a name and location.

2. Create the folder structure:

cd ~/vault  # adjust to your path
mkdir -p permanent inbox fleeting templates logs references
mkdir -p my-project/{architecture,pipeline,data,features,logs}

3. Create the CLAUDE.md:

This is the file Claude Code reads automatically. Create CLAUDE.md at the vault root:

# Vault — Instructions for Claude Code

## What is this vault
Centralized knowledge base for all projects.
Persistent memory across sessions.

## Project stacks
- Project X: React + Supabase
- Project Y: Python + FastAPI
(adapt to your projects)

## Zettelkasten Rules

### Note creation
- Use wikilinks: [[note-name]] (not markdown links)
- Mandatory YAML frontmatter on every note
- Filenames in kebab-case: `auth-flow.md`, not `Auth Flow.md`
- 1 concept per permanent note (atomicity)
- Minimum 2 wikilinks per note (dense linking)

### Standard frontmatter
---
title: Note Name
tags: [project, topic]
created: YYYY-MM-DD
updated: YYYY-MM-DD
status: active
type: permanent
---

### Never do
- Don't delete notes without asking
- Don't use markdown links for internal notes (use wikilinks)
- Don't create notes without frontmatter
- Don't change folder structure without documenting it

## Session Commands

### /resume
When you receive this command:
1. Read the 3 most recent session logs in logs/
2. Read architecture/decisions.md for the current project
3. Summarize current state and what's left to do

### /save
When you receive this command:
1. Create a session log in logs/YYYY-MM-DD-description.md
2. Record: what was done, decisions made, pending items
3. Add wikilinks to created/modified notes
4. Run git commit + push if in a repository

4. Create a note template:

cat > templates/default-note.md << 'EOF'
---
title: {{title}}
tags: []
created: {{date}}
updated: {{date}}
status: draft
type: permanent
---

# {{title}}

## Context

## Details

## Related links
EOF

5. Recommended Obsidian plugins:

Part 2 — Chat Import Pipeline

Concept

Your Claude chats (both Code and Web) contain valuable decisions, insights, and context that get lost in the history. This pipeline exports, processes, and imports those conversations as vault notes — with frontmatter, automatic tags, and wikilinks to existing notes.

Components

~/scripts/
├── claude_to_obsidian.py          # processor (frontmatter, tags, wikilinks)
└── sync_claude_obsidian.sh        # automation (export + process)

~/claude-exports/                   # temporary staging area (outside vault)
├── code/                           # Claude Code exports
└── web/                            # Claude Web exports

Setup

1. Install the Claude Code extractor:

pip install claude-conversation-extractor

2. Create staging directories:

mkdir -p ~/claude-exports/code ~/claude-exports/web

3. Create the post-processing script (~/scripts/claude_to_obsidian.py):

The script should:

• Read each exported .md file
• Detect origin (Code vs Web)
• Generate automatic tags based on content keywords
• Add standardized YAML frontmatter
• Insert [[wikilinks]] for notes that already exist in the vault
• Copy to chats/code/ or chats/web/ inside the vault

Example keyword-to-tag mapping:

KEYWORD_TAG_MAP = {
    "python": "python",
    "react": "react",
    "supabase": "supabase",
    "deploy": "deploy",
    "bug": "debugging",
    "refactor": "refactoring",
    # add your own
}

4. Create the automation script (~/scripts/sync_claude_obsidian.sh):

#!/bin/bash
EXPORT_DIR="$HOME/claude-exports"
VAULT_DIR="$HOME/vault"  # adjust to your path
SCRIPT_DIR="$HOME/scripts"
LOG="$SCRIPT_DIR/sync.log"

echo "[$(date)] Sync started" >> "$LOG"

# Export Claude Code chats
claude-extract --all --output "$EXPORT_DIR/code" 2>> "$LOG"

# Process and send to vault
python3 "$SCRIPT_DIR/claude_to_obsidian.py" \
    --export-dir "$EXPORT_DIR" \
    --vault-dir "$VAULT_DIR" \
    --move 2>> "$LOG"

echo "[$(date)] Sync completed" >> "$LOG"

5. Schedule automatic execution:

chmod +x ~/scripts/sync_claude_obsidian.sh

# Run daily at 10 PM
(crontab -l 2>/dev/null; echo "0 22 * * * $HOME/scripts/sync_claude_obsidian.sh") | crontab -

6. For Claude Web chats:

Install the "Export Claude Chat to Markdown" browser extension for Chrome/Edge. Do periodic bulk exports, save the .md files to ~/claude-exports/web/, and the cron job handles the rest.

7. Add a section to the vault's CLAUDE.md:

## Chat Import Pipeline

### Structure
- `chats/code/` → imported Claude Code conversations
- `chats/web/` → imported Claude Web/App conversations
- All chats get frontmatter wit