页面 · CaBrain

CaBrain brains — FlowOS & AVO

CaBrain hosts multiple namespaces ("brains"):

flowos — the FlowOS / OneStudio hub (below).
avo (~2,490 memories) — the AVO "Founder Readiness Lab" repo (onestudio-exp/AVO) markdown: 359 files chunked by heading (board pack, playbooks, kaizen engine, kickstart legal/finance, pitch/fundraising, research, theses, drills). Re-run with scratchpad/avoingest.py (needs the repo cloned to scratchpad/AVO via gh repo clone onestudio-exp/AVO). Query with memory_recall(namespace="avo", query="…").

The FlowOS brain — querying CaBrain about ventures & people

CaBrain has ingested the FlowOS (OneStudio hub) knowledge into the flowos namespace so any MCP-connected session can ask about a venture, agent, person, learning, or lean canvas and get the answer back from memory.

What's ingested (namespace `flowos`, ~1,780 memories)

Two passes: the FlowOS MCP (structured API) and the full hub Postgres DB (onestudio_hub, via an SSH tunnel) — everything except high-frequency telemetry (feature/transition/claude-activity events, github metrics, views, tokens).

Type	Source
Ventures (66, full records + team)	DB `ventures` + MCP `get_venture`
Domain-expert agents (40, en+ar, skills)	DB `agents` + MCP `get_agent`
People (46, names/emails/roles + venture memberships)	DB `users` + `venture_members`
Issues (597 — title, description, status, priority)	DB `issues`
Posts / status feed (930)	DB `posts`
Learnings (178)	MCP `list_learnings` + DB
Lean canvases (57)	MCP `get_lean_canvas`
Roadmaps (178), Releases (69), Goals (222), Harvested research (82), Tasks (68)	DB

Near-duplicate/templated rows collapse via the §4.1 write-decision (that's why the row total is ~1,780, not the raw ~2,600). Each memory embeds via TEI (bge-m3) and is recallable by vector + BM25 + rerank. Re-run / extend with scratchpad/ingest.py, enrich.py, and dbingest.py (the DB pass needs the SSH tunnel: ssh -L 15432:<DB_HOST>:5432 <USER>@<JUMP_HOST> — real host/creds held by the owner).

How to query it (from a new Claude Code session)

The cabrain MCP server is wired in .mcp.json (stdio, brain-mcp, talks to the app on localhost:8080). After restarting Claude Code so it loads, the agent has these tools: memory_recall, memory_retain, memory_get, memory_forget, memory_share, memory_recall_archive.

Ask with memory_recall, namespace flowos:

text

Verified live: "Sentra" → Sentra, "Zeedly embedded Salla app" → Zeedly, "PDPL Saudi data protection" → PDPL Starter Kit, "football scouting in Saudi Arabia" → Akhdar, "permission-check learning" → the exact learning.

Running / restarting the app

The app must be up for the brain MCP to reach it. It runs detached:

bash

run-cabrain.sh sets DATABASE_URL (from CABRAIN_DATABASE_URL), the TEI/Cognee URLs, CACHE_DRIVER=redis, and BRAIN_BM25_TOKENIZER. The workspace must be on stack_stacknet (sudo docker network connect stack_stacknet coder-… on the WSL host) so pg/tei-embed/cognee/redis resolve.

Optionally add the live FlowOS MCP (not committed — has a token)

To also give the session live FlowOS hub tools, add this to a non-committed config (~/.claude.json user scope), NOT the repo .mcp.json:

jsonc

Two follow-ups that sharpen the brain (infra)

Production BM25 tokenizer — name/keyword recall is strongest with a real multilingual tokenizer. A superuser runs infra/grant-bm25.sql §3 to create cabrain_ml (llmlingua2), then set BRAIN_BM25_TOKENIZER=cabrain_ml.
Cognee graph — cognify currently fails with "Missing required pgvector credentials" (Cognee's own vector store isn't configured). Once infra sets Cognee's pgvector creds, brainctl mirror flowos populates the entity graph + Graph Explorer.