Local-first 🧠memory for Hermes Agent that lives in your Obsidian vault and remembers project context. Nightly 😴 dream passes turn repeat corrections into confirmed preferences with measurable confidence. Adapters ship for Claude Code, Codex, and OpenClaw, with an MCP server for anything else.
Grok Build
o2b install --target grok --apply performs a native Grok Build
integration in one pass, writing to grok's own primary config locations:
- MCP servers - registers
open-second-brainandopen-second-brain-writerin${GROK_HOME:-$HOME/.grok}/config.tomlunder[mcp_servers.*](grok's highest-priority MCP source), each with an absolute command (bun run <repo>/src/cli/main.ts mcp …). User-authored[mcp_servers.*]and other config sections are preserved. - Lifecycle hooks - writes
${GROK_HOME:-$HOME/.grok}/hooks/open-second-brain.json(grok's native, always-trusted hooks dir), again with absolutebuncommands.
Why an absolute command rather than a bare o2b: verified against live grok
0.2.45 (the session debug log), grok spawns MCP servers and hook scripts with a
restricted PATH that excludes ~/.local/bin, so a bare o2b fails to spawn and
no tools load. The absolute bun run …/main.ts form is what grok actually
launches and handshakes in a session (71 + 5 tools registered, hooks dispatched).
Install
o2b install --target grok --apply --yes
Start grok (or press r in the /mcps and /hooks modals) to load them.
Confirm with grok inspect and, for a live check, grok mcp doctor open-second-brain.
What it provides
- MCP servers -
open-second-brain(full Brain surface) andopen-second-brain-writer(the always-loaded writer set). grok namespaces the tools asopen-second-brain__<tool>/open-second-brain-writer__<tool>and the model reaches them through grok'ssearch_tool/use_tool. - Active context inject -
SessionStartrunshooks/active-inject.ts, appending the renderedBrain/active.mddigest so the agent sees live preferences without callingbrain_queryfirst. - Post-write reminder - after a file-mutating tool (grok's
search_replace, plus the Claude-styleWrite/Edit/MultiEdit/apply_patchaliases)hooks/post-write-reminder.tsappends the logging nudge. - Session capture and the stop-log guardrail - the
Stop/SessionEndhooks capture the session and check that an artifact turn logged a Brain event, mirroring the Claude Code behavior.
Every hook is fail-soft: a missing vault or runtime error never breaks the grok session.
grok's Brain writes (hook captures and tool calls) attribute to a grok-specific
identity derived from your configured agent_name: the host segment is kept and
the vendor token is swapped to grok (claude-vps-agent -> grok-vps-agent; a
name that does not fit the <vendor>-<host>-agent shape is prefixed with grok-).
So grok activity is distinguishable per runtime - and, in a shared multi-device
vault, per device - rather than logged under the shared identity. Both the MCP
env (config.toml) and the hooks file carry this same derived name.
Importing grok sessions
grok stores each session as
${GROK_HOME:-$HOME/.grok}/sessions/<encoded-cwd>/<id>/updates.jsonl (an ACP
session-update stream). Import one session, or a directory of them, into the
Brain:
o2b brain import-session \
~/.grok/sessions/<encoded-cwd>/<session-id>/updates.jsonl \
--vault /path/to/vault
Autodetect resolves the grok format; pass --format grok to force it. The
importer extracts @osb markers from messages and replays brain_feedback
tool calls (grok's open-second-brain__brain_feedback is normalized to the
bare name first).
Verify
o2b install --check --target grok
Reports drift when the config.toml MCP tables or the hooks file differ from
canonical (for example after an Open Second Brain upgrade, or if bun/the repo
moved - re-run apply to refresh the absolute paths).
A successful check prints:
o2b install --check
--------------------
grok ok
This block is asserted against the adapter's real verify() output
by tests/docs/install-verify-conformance.test.ts, so it cannot
drift from the code.
Uninstall
o2b uninstall --target grok --apply
Removes the two [mcp_servers.*] tables from config.toml and deletes the
hooks file. Unrelated grok configuration (other MCP servers, [cli],
[marketplace]) is left intact.
Relationship to the Claude Code integration
Grok reads Claude Code configuration for compatibility (MCP from
~/.claude.json, hooks from ~/.claude/settings.json). An operator who already
runs the Open Second Brain Claude Code integration may see some of it picked up
that way. The native grok target exists so that operators who run only grok
get a first-class, in-session-working install, and so grok sessions import
into the Brain (no compatibility path covers that direction). The
integration uses grok's own config.toml and hooks dir - it does not write into
the ~/.claude/ namespace.
Notes
- Verified against grok 0.2.45 and its bundled docs (
~/.grok/docs/user-guide/). GROK_HOMEoverrides the~/.grokbase directory; the adapter honors it.- The absolute
bun/repo paths are resolved at install time. If you move the Open Second Brain checkout or change thebunbinary, re-runo2b install --target grok --apply --yes. - Rules: grok reads
AGENTS.mdandCLAUDE.mdnatively, so vault-level rules can travel as project instructions without any Open Second Brain involvement.