Claude Memory Visualizer

3D visualization of embedding spaces from claude-memory - explore your Claude Code conversations in semantic space.

Prerequisites

This visualizer works with data from claude-memory. You'll need:

1. Bun - Fast JavaScript runtime

   curl -fsSL https://bun.sh/install | bash

2. claude-memory installed and synced - Follow the claude-memory setup guide

3. Python dependencies for the export script (or use claude-memory's venv):

   pip install chromadb umap-learn scikit-learn numpy
   # Or if you have claude-memory installed:
   # ~/dev/claude-memory/.venv/bin/python scripts/export-chromadb.py

4. Ollama (optional) - For AI summaries when box-selecting points:

   ollama pull qwen2.5:1.5b   # or any text model

Quick Start

# Clone the repo
git clone https://github.com/jopnelli/claude-memory-visualizer
cd claude-memory-visualizer

# Install dependencies
bun install

# Export your claude-memory data (skip if just trying the demo)
python scripts/export-chromadb.py

# Start dev server
bun run dev

Open http://localhost:5173

The visualizer auto-loads your Claude Memory data if available, otherwise shows a demo dataset.

Features

• 3D Point Cloud - Visualize thousands of conversation embeddings in 3D space
• Pre-computed Projections - UMAP, t-SNE, PCA computed server-side for instant switching
• Semantic Search - Find conversations by meaning using all-mpnet-base-v2 embeddings
• Box Selection - Cmd+drag to select regions, get AI summaries via Ollama
• Time-based Coloring - Older conversations in blue, newer in red
• Inspector Panel - Hover/click to see full document text and metadata

Export Script

The Python export script (scripts/export-chromadb.py) reads from ChromaDB and pre-computes all three projections:

# Basic usage (reads from ~/.claude-memory/chroma)
python scripts/export-chromadb.py

# Custom options
python scripts/export-chromadb.py \
  --chroma-path /path/to/chroma \
  --output public/data/my-data.json \
  --limit 1000

Output format:

{
  "metadata": {
    "name": "Claude Memory",
    "embedding_model": "sentence-transformers/all-mpnet-base-v2",
    "embedding_dim": 768,
    "count": 3486
  },
  "documents": [...],
  "projections": {
    "umap": [[x, y, z], ...],
    "tsne": [[x, y, z], ...],
    "pca": [[x, y, z], ...]
  }
}

Data Sources

1. Claude Memory - Your exported claude-memory data with full semantic search
2. Demo Dataset - A small sample dataset to explore the interface (text search only)

Tech Stack

• Three.js - 3D rendering with point clouds and OrbitControls
• DRUIDJS - Browser-side dimensionality reduction (fallback if no pre-computed projections)
• Transformers.js - In-browser all-mpnet-base-v2 embeddings for semantic search (~420MB, cached)
• Ollama - Local AI summaries for box selection (optional)
• Bun + Vite - Dev server and bundling
• TypeScript - Type-safe codebase

Algorithms

Algorithm	Speed	Quality	Best For
UMAP	Fast	High	Default choice, good balance
t-SNE	Slow	High	Tight, well-separated clusters
PCA	Instant	Medium	Quick overview, may have overlap

Known Limitations

• Turn-level chunking: claude-memory chunks at individual turns (1 user + 1 assistant message), not full conversations. Search may miss context that spans multiple turns.
• First search delay: Downloads ~420MB embedding model on first search (cached after)
• Best for <10k docs: Larger datasets may be slow to render

Development

# Start dev server
bun run dev

# Re-export data after claude-memory changes
python scripts/export-chromadb.py

Faster Semantic Search (Optional)

By default, semantic search downloads a ~420MB embedding model to the browser on first use. For faster search, run the local embedding server:

pip install sentence-transformers
python scripts/embed-server.py

Tip: If you have claude-memory installed, its venv already includes sentence-transformers.

The visualizer auto-detects the local server and uses it for faster search.

License

MIT