Architecture diagrams
ProjectBrain — portfolio intelligence for agents
Local SQLite + vectors + graph: ingest charters/decisions, score portfolio health, MCP tools for IDE session handoff, playbooks, and semantic search. Cloud automations stay repo-only.
Canonical architecture for Memory. Linked from the project essay; not listed in Blog.
ProjectBrain — system overview
Personal project intelligence for a multi-repo Cursor portfolio. Git remains source of truth per repo; ProjectBrain ingests charters, decisions, and reflections into a local store agents query through MCP.
Ingest and query loop
Runtime boundaries
| Runtime | May use ProjectBrain MCP? | Memory source |
| --- | --- | --- |
| Cursor IDE (local) | Yes | MCP + repo handoff files |
| Cursor Automations (cloud) | No | Target repo .state/ / playbooks on GitHub only |
Milestone 4 capabilities
- Installer,
doctor, backup/restore, encrypted export sync - Scoring (success, risk, momentum, value, fulfillment, technical debt)
- Playbook extraction from winner projects;
setup_new_projectone-shot kit - Semantic search (LanceDB), graph traverse (Kùzu), portfolio summary
- Evolution scout + adoption eval gate; monthly digest
Data location
Default: %USERPROFILE%\.projectbrain\ or repo-local .projectbrain/ when PROJECTBRAIN_DATA_DIR is set.