🧠 LocalBrain

Chat with your entire AI history.

Index conversations from Pi, Claude Code, Codex, DROID/Factory, Qwen Code, Hermes, OpenClaw, Gemini & more — then query them with a local LLM. Zero cloud. All yours.

Quick Start · Sources · Models · Architecture · Extract Data

---

✨ What Is This?

LocalBrain is a privacy-first local AI assistant that indexes all your AI conversations — from coding sessions (Pi, Claude, Codex, Hermes, OpenClaw, Factory/DROID, Qwen Code), chat takeouts (Gemini), and notes — then lets you chat with your entire history using GPU-accelerated local inference.

Think of it as a personal second brain that actually remembers everything.

✨ Features

Feature	Details
10+ sources	Pi, Claude, Codex, OpenCode, Cursor, Factory, Hermes, OpenClaw, Qwen Code, Gemini
Hybrid retrieval	BM25 keyword search + cross-encoder reranking (ms-marco-MiniLM-L-6-v2)
Local LLM	Qwen 3.5, Gemma 4, Llama 3.3, Mistral — any GGUF via llama.cpp
Web UI	Clean chat interface with source citations
Personal profile	Edit runtime/personal.json for contextual answers about you
100% private	No cloud APIs, no data leaves your machine, no vector DB, MIT licensed

🚀 Quick Start

# 1. Install dependencies
pip install -r requirements.txt

# 2. Initialize database
python scripts/ops/init_db.py

# 3. Place your data exports in exports/
#    (see Data Extraction section below)

# 4. Run full ingestion pipeline
python scripts/ingest/discover.py
python scripts/ingest/classify.py
python scripts/ingest/extract.py
python scripts/ingest/chunk.py

# 5. Start llama-server with a GGUF model
llama-server -m models/qwen3.5-4b.gguf --port 8080 -ngl 99 -c 8192

# 6. Start web UI -> open http://localhost:8096
python scripts/web_server.py

# Or query from CLI
python scripts/query.py "What have I worked on?" --db data/state.db

📦 Supported Sources

Source	Extraction	Status
Pi	pi-brain export pi	✅
Claude Code	pi-brain export claude	✅
Codex	pi-brain export codex	✅
OpenCode	Auto-discovered	✅
Cursor	Pre-extracted JSONL	✅
Factory / DROID	Auto-discovered	✅
Hermes	pi-brain export hermes	✅
OpenClaw	pi-brain export openclaw	✅
Qwen Code	Auto-discovered	✅
Google Gemini	Takeout HTML parser	✅

🤖 Recommended Models

Model	Size (Q4_K_M)	Min VRAM	Best For
Gemma 4 12B	~8 GB	10 GB	⭐ Best overall
Qwen 3.5 4B	~3 GB	4 GB	🚀 Lightest / testing
Qwen 3.5 9B	~6 GB	8 GB	🎯 Sweet spot
Qwen 3.5 27B	~17 GB	20 GB	🧠 Best quality
Qwen 3.5 60B	~36 GB	40 GB	🔬 Maximum reasoning
Gemma 4 48B	~30 GB	36 GB	✍️ Creative writing
Llama 3.3 8B	~5 GB	6 GB	🌐 General purpose
Mistral Small 3.1	~14 GB	16 GB	💻 Coding

Quick pick: Qwen 3.5 4B for testing, Gemma 4 12B for daily use, Qwen 3.5 27B for best quality. Download GGUF quants from Hugging Face.

CPU-Only Mode

No GPU? llama-server runs on CPU too — set gpu_layers: 0 in config/models.yaml. Expect ~10-20 tok/s on a modern CPU with 8B models.

Apple Silicon

On Mac with M-series chips, llama.cpp auto-detects Metal GPU. Set gpu_layers: 99 for full offload.

🏗 Architecture

+--------------------------------------------------------------+
|                        LOCALBRAIN                             |
+--------------------------------------------------------------+
|  DATA SOURCES                                                 |
|  Pi . Claude . Codex . OpenCode . Cursor . Factory            |
|  Hermes . OpenClaw . Qwen Code . Gemini Takeout               |
+---------------------------+----------------------------------+
|                           v                                  |
|  INGESTION PIPELINE                                           |
|  discover -> classify -> extract -> chunk -> index            |
+---------------------------+----------------------------------+
|                           v                                  |
|  SQLite (content_chunks) + BM25 in-memory index               |
+---------------+-------------------+--------------------------+
|               v                   v                          |
|  +-----------+-----+     +--------+-----------+              |
|  |  Web UI (:8096) |     |  CLI Query         |              |
|  +-----------+-----+     +--------+-----------+              |
|               |                   |                          |
|               +--------+----------+                          |
|                        v                                     |
|  +----------------------------------------------------+      |
|  |  HYBRID RETRIEVAL                                    |      |
|  |  1. BM25 -> top 50 candidates (keyword match)       |      |
|  |  2. Cross-encoder -> rerank (semantic scoring)      |      |
|  |  3. Return top 10                                   |      |
|  +----------------------------------------------------+      |
|                        v                                     |
|  +----------------------------------------------------+      |
|  |  llama-server (:8080)                               |      |
|  |  Qwen 3.5 / Gemma 4 / Llama 3.3 / Any GGUF          |      |
|  |  Full GPU offload . Apple Silicon . CPU fallback      |      |
|  +----------------------------------------------------+      |
+--------------------------------------------------------------+

Why Hybrid Retrieval (Not Vector Embeddings)?

Approach	Storage	Speed	Accuracy	Complexity
BM25 only	None	Fastest	Good	Simple
Hybrid (BM25 + reranker)	None	Fast	Best	Moderate
Vector embeddings	622K x 768d (~2GB)	Medium	Good	Complex

Hybrid retrieval is what production RAG systems actually use. The cross-encoder sees the query and document together (not as independent embeddings), giving much more accurate relevance scores.

📂 Data Extraction

From pi-brain (Pi, Claude, Codex, OpenCode, Cursor, Hermes, OpenClaw)

npm install -g @0xsero/pi-brain

pi-brain export pi --output exports/pi_sessions.jsonl
pi-brain export claude --output exports/claude_sessions.jsonl
pi-brain export codex --output exports/codex_sessions.jsonl
pi-brain export hermes --output exports/hermes_sessions.jsonl

Enhanced pi-brain fork with Factory/DROID, Gemini Takeout & Qwen Code plugins: github.com/0xSero/pi-brain

From Factory / DROID

Auto-discovered from ~/.factory/sessions/:

python scripts/ingest/factory_extractor.py

From Qwen Code

Auto-discovered from ~/.qwen/projects/*/chats/:

python scripts/ingest/qwen_extractor.py

From Google Gemini Takeout

1. Download from Google Takeout -> select "Gemini"
2. Place ZIP(s) in exports/
3. Extract and normalize:

python parse_gemini_html.py    # Streaming parser for 60MB+ HTML files
python normalize_gemini.py      # Converts to ingestable JSONL

⚙ Configuration

Personal Profile

Edit runtime/personal.json so the AI knows about you:

{
  "name": "Your Name",
  "location": "Your Location",
  "hardware": { "gpu": "Your GPU", "ram_system": "Your RAM" },
  "tools_used": ["Pi", "Claude Code", "Codex", "Gemini"],
  "projects": ["Project 1", "Project 2"],
  "interests": ["AI", "coding"]
}

Model Settings

Edit config/models.yaml for model paths, GPU layers, and context size.

🧪 Benchmark

python benchmark.py

Runs 20 diverse questions scoring retrieval accuracy + LLM answer quality. Typical score: 94% (75/80).

📁 Project Structure

localbrain/
├── config/             # YAML configs (models, sources)
├── data/               # SQLite database
├── exports/            # Raw data sources (takeout ZIPs, JSONL)
├── runtime/            # Web UI (index.html, app.js, style.css)
├── scripts/
│   ├── core/           # Hybrid search (BM25 + cross-encoder), LLM client
│   ├── ingest/         # Pipeline (discover->classify->extract->chunk)
│   └── ops/            # Database operations
├── models/             # GGUF model files (download separately)
├── parse_gemini_html.py  # Google Takeout HTML parser
└── benchmark.py          # Quality test suite

🔒 Privacy Guarantees

• 100% local — no data ever leaves your machine
• No cloud APIs — inference runs via llama.cpp on your hardware
• No vector database — BM25 + cross-encoder, zero storage overhead
• SQLite only — no external databases or services
• Open source — audit everything, MIT licensed

🤝 Related Projects

Project	Description
pi-brain	Original session extractor — Pi, Claude, Codex exports
SergiioB/pi-brain	Enhanced fork with Factory/DROID, Hermes, OpenClaw, Gemini Takeout & Qwen Code plugins

📜 License

MIT — do whatever you want with it.