Cerebro is a local-first storage system. All data resides in SQLite files on the local filesystem. There are no network listeners, no servers, and no remote data transmission in the core library.

• SQLite with WAL mode — Write-Ahead Logging for crash safety and concurrent read access
• Single-writer model — one process writes to a given brain at a time, enforced by SQLite's locking
• File permissions — brain files are created with user-only permissions; the calling application is responsible for directory permissions

Cerebro supports external embedding providers (Voyage AI, Ollama). API keys for these providers:

• Are passed via brain.EmbedConfig at initialization time
• Are stored in the brain's metadata table (SQLite) for reopening
• Are not logged or included in export bundles
• For CLI usage, should be set via environment variables (e.g., CEREBRO_VOYAGE_API_KEY)

• cerebro export outputs the full brain contents (nodes, edges, metadata) as JSON, SQL, or SQLite copy
• Exported data may contain sensitive information stored in memory nodes
• Import uses conflict resolution (skip or replace) and does not execute any embedded content

• Does not make network requests (except via embedding providers when configured)
• Does not execute code from memory content
• Does not evaluate or interpret memory node content — it is treated as opaque text
• Does not manage authentication or authorization — that is the responsibility of the consuming application

If you discover a security vulnerability in Cerebro, please report it responsibly:

1. Do not open a public GitHub issue
2. Email the maintainer directly at the address listed in the repository profile
3. Include a description of the vulnerability, steps to reproduce, and potential impact
4. Allow reasonable time for a fix before public disclosure

We aim to acknowledge reports within 48 hours and release fixes promptly.