Obsidian vault as shared memory for Claude Code, Codex, Cursor and Grok Bot

What they did

They took the obsidian-second-brain repo, which ships 8 agents and 48 commands, and pointed several different coding CLIs at the same Obsidian vault so the vault, not the CLI, holds the memory. Claude Code, Codex, Cursor and Grok Bot all read and write the same notes, and the Crew layer adds a plain-folder logbook so each agent can see what the others finished and leave a lesson behind. Lessons only become rules when a human runs `keep`, so the agents propose but the person promotes. The point was continuity: switching CLIs no longer means losing context, because the vault outlives whichever tool is in fashion.

How it went

The setup gives a genuinely portable second brain: notes, decisions and job records survive a CLI swap, and a new session can answer what another agent did earlier without re-explaining. Polish is high for a free MIT project, with install paths documented for Claude Code, Codex, Gemini, OpenCode, Antigravity, Hermes, Pi and Grok Bot. Limits: it is a workflow scaffold, not a hosted service, so quality depends on how disciplined the vault is kept.

Worth knowing

Crew runs no background process, so nothing syncs unless an agent session is active. The human `keep` gate means lessons pile up for review rather than self-applying. You need an Obsidian vault (or a plain folder for Crew) plus at least one supported CLI installed and authenticated, and multi-agent use assumes everyone points at the same path.

Wiring

Install the repo per its README for your CLI, then set the vault path so every agent resolves the same directory. Obsidian-side: open the vault, enable the shipped agents and commands. Crew-side: create a plain logbook folder (Obsidian not required) that Claude Code, Codex, Cursor and Grok Bot all write job records into; each session reads prior records on start. Lessons are written as proposals and only promoted to rules by a human running `keep`. No daemon runs in the background, so ordering is session-driven: agent A finishes and logs, agent B starts and reads.

Try it yourself

Set up obsidian-second-brain in this repo. Create or point to an Obsidian vault, install the agents and commands for my CLI (Claude Code), and configure the Crew logbook folder so Codex and Cursor sessions can read the same job records. Then run one small task end to end, leave a lesson, and show me the log entry a second agent would see.