Extract what your browser knows about you into a self-ranking SQLite database. Reads autofill, login data, browsing history, bookmarks, WhatsApp contacts, LinkedIn connections, and Notion workspaces — directly from local browser files.
| Source | Data | Browser files |
|---|---|---|
| Web Data | Autofill, addresses, credit cards | Web Data SQLite |
| Login Data | Accounts, emails, usernames | Login Data SQLite |
| History | Tool/service usage frequency | History SQLite |
| Bookmarks | Interests, saved tools | Bookmarks JSON |
| IndexedDB | WhatsApp contacts | LevelDB via ccl_chromium_reader |
| Local Storage | LinkedIn connections | LevelDB via ccl_chromium_reader |
| Notion | Workspace users, pages | IndexedDB |
Supported browsers: Arc, Chrome, Brave, Edge, Safari, Firefox.
npx ai-browser-profile init # sets up ~/ai-browser-profile, Python venv, core deps
npx ai-browser-profile install-embeddings # optional: semantic search (~180MB)Requires Python 3.10+ and Node.js 16+. macOS only (reads from ~/Library/Application Support/).
This creates ~/ai-browser-profile/ with a Python venv, installs dependencies, and symlinks Claude Code skills to ~/.claude/skills/.
cd ~/ai-browser-profile && source .venv/bin/activate
python extract.py # scan all browsers
python extract.py --browsers arc chrome # specific browsers
python extract.py --no-indexeddb --no-localstorage # skip LevelDB (faster)
python extract.py --output /path/to/memories.db # custom output pathTo update after a new release:
npx ai-browser-profile update # updates code, preserves memories.dbfrom ai_browser_profile import MemoryDB, extract_memories
# Extract from browsers
mem = extract_memories("memories.db")
# Query
mem.search(tags=["identity", "contact_info"])
mem.text_search("github")
mem.semantic_search("what tools do I use most")
# Profile summary
print(mem.profile_text())
# History + supersession chain
mem.history("email")Self-ranking — each memory tracks appeared_count (how often it was seen during extraction) and accessed_count (how often it was queried). The ratio accessed_count / appeared_count is the hit_rate, used to surface the most relevant memories.
Semantic dedup — new entries are compared against existing ones using nomic-embed-text-v1.5 embeddings (768-dim, ONNX Runtime). If cosine similarity >= 0.92 with the same key prefix, the old entry is superseded rather than duplicated.
Key schema — memories use structured keys with cardinality rules:
- Single-value (
first_name,last_name,full_name, ...): new values automatically supersede old ones - Multi-value (
email,phone,account:github.com,tool:vscode, ...): multiple values coexist
Entity linking — accounts sharing the same username/email are automatically linked via same_identity relations.
memories (id, key, value, confidence, source, appeared_count, accessed_count,
created_at, last_appeared_at, last_accessed_at, superseded_by,
superseded_at, search_text, reviewed_at)
memory_tags (memory_id, tag) -- identity, contact_info, address, payment,
-- account, tool, contact, work, knowledge,
-- communication, social, finance
memory_links (source_id, target_id, relation, created_at)
memory_embeddings (memory_id, embedding) -- 768-dim BLOBextract.py # CLI entry point
ai_browser_profile/
__init__.py # exports MemoryDB, extract_memories
db.py # MemoryDB: schema, upsert, search, profile
embeddings.py # ONNX Runtime embeddings + cosine search
extract.py # extraction orchestrator
ingestors/
browser_detect.py # find browser profiles
constants.py # lookup maps, browser paths
webdata.py # autofill, addresses, credit cards
history.py # browsing history → tool usage
logins.py # saved logins → accounts
bookmarks.py # bookmarks → interests
indexeddb.py # WhatsApp contacts
localstorage.py # LinkedIn connections
notion.py # Notion workspace data
messages.py # message extraction
MIT