joshu-brain

Call the gbrain query MCP tool with your question, source_id set to "__all__" for cross-source search, a limit of 10-20, and recency enabled with a window like 90d. Do not rely on the search tool alone, since it only covers the default source.

The search tool is keyword full-text search limited to the default source (joshu's files and journals), while query performs semantic cross-source retrieval. Query is required for federated paths like investors or workspace folders where search often returns empty.

Use Hindsight memory for conversation history, preferences, and facts from past chats. Use gbrain when the answer lives in documents such as journals, research notes, uploads, or PDF extracts saved as files.

All user-visible files must be written under JOSHU_FILES_ROOT, the joshu's files folder on the ArozOS Desktop. Never write to macOS ~/Desktop, the Desktop root, or paths outside JOSHU_FILES_ROOT.

Folders outside joshu's files live in a federated source, so the default-source search tool misses them. Use query with source_id "__all__", and fall back to chunk_text from results when get_page returns page_not_found for a federated slug.

Joshu automatically extracts a sibling markdown file next to any PDF or .txt dropped on the Desktop, and gbrain indexes that markdown. Scanned image PDFs may fail extraction and need OCR or manual markdown.