Brain — persistent memory for your AI agents
Your agents forget everything between sessions. Brain doesn't.
Brain is a hosted knowledge base your AI agents read and write overMCP — architecture decisions, runbooks,hard-won gotchas, and agent-to-agent handoffs, searchable from Claude Code,Codex, and any MCP-compatible client. Stop re-explaining your stack to everynew session.
Install (2 minutes)
Claude Code:
claude plugin marketplace add levirge/brain
claude plugin install brain@levirge-brain
Then run /mcp in a session and sign in.
Codex CLI:
codex plugin add levirge/brain
Cline, or any MCP client — remote server, no install:
{ "mcpServers": { "brain": { "type": "streamableHttp", "url": "https://brain.levirge.com/mcp" } } }
That is Cline's spelling of the transport type; other clients name it differently. OAuthsign-in on first connect. Details: llms-install.md.
What you get
- Search-first memory —
search_knowledgewith semantic search + reranking;agents cite what they read instead of hallucinating. - Write-back that compounds —
add_knowledgecaptures decisions and fixesonce; every future session starts warmer. - Fact lineage — trace any answer back to the documents and atoms it camefrom (
/brain:lineage). - Agent-to-agent handoffs — a shared inbox between agents and machines:hand work from one session, role, or CLI to another (
/brain:handoff). - Work sessions & context packs —
/brain:startpulls the rightcontext for a task;/brain:finishpersists what was learned. - Self-improving ranking —
feedback(helpful / dead-end verdicts) tunesretrieval to how your team actually works.
Works with your models
Measured in a 10-task × 3-repeat bake-off (fixed judge): Claude Opus-classmodels ground perfectly, and the bundled search-first contract doubledretrieval compliance on a local 26B. Details inplugins/brain/README.md.
Links
- App / sign in: https://brain.levirge.com
- Levirge: https://levirge.com
- Plugin reference: plugins/brain/README.md
Plugin versioning is semver, independent of the Brain server release.