second-brain

One memory for Claude, Codex, Gemini, OpenCode, and Cline — stored as files you own.

second-brain is a local-first knowledge graph for AI agents. Its canonical store is an OKF-flavoured Markdown Bundle; SQLite is a disposable search index. The same memory is available through an Agent Skill, a command-line interface, and a stdio MCP server.

Beta, with evidence. The package, local engine, MCP subprocess, Obsidian format behaviour, snapshot/restore path, and 262-test suite are checked in. Live activation in every agent host and live cloud-provider accounts are not yet part of CI. See the compatibility matrix.

Why this exists

Ask one agent to remember an architectural decision, switch tools, and the next agent usually starts cold. Hosted memory products solve that by putting another vendor between you and your own history.

second-brain takes the opposite approach:

• Files are truth. Notes remain inspectable Markdown with stable IDs and

wikilinks.

• Agents are clients. Claude, Codex, Gemini, OpenCode, and Cline share the

same open skill package; MCP-capable clients can call the same brain over JSON-RPC. Host activation evidence is tracked explicitly below.

• Indexes are replaceable. Delete SQLite, rebuild it from the Bundle, and

continue.

• Writes are paired. Supported CLI/MCP mutations flush Markdown before they

report success; a lock, receipt, and crash journal refuse stale or mixed generations instead of guessing.