🧠 Company Brain

One brain for every agent in your company.

Local-first memory engine that unifies Microsoft Fabric IQ, Work IQ, and Foundry IQ into a single Postgres + pgvector store — and serves it to any agent over one API.

Quickstart · Architecture · Connectors · API · Roadmap · FAQ

---

✨ Why Company Brain

Microsoft's "IQ" surfaces are powerful, but siloed:

Surface	What it knows	What it can't do
Fabric IQ	Your data warehouse, lakehouses, semantic models	Can't see your inbox
Work IQ (Copilot/Graph)	Your mail, Teams, SharePoint, calendar	Can't see your data ontology
Foundry IQ	Your AI project knowledge bases	Can't see either of the above

None of them talk to each other. None of them give you raw access to facts, embeddings, or a unified retrieval layer. Your agents end up with three half-brains, none of them yours.

Company Brain is the missing piece: a small, opinionated MemoryStore interface that ingests all three sources, extracts atomic facts, stores them in your own Postgres, and serves them to any agent — over one HTTP API, on your machine, with your keys.

Shared brain first. Boundaries second. Your data stays under your tenant. Your agents share one memory. Permissions are enforced by container_tag at query time, not by data duplication.

---

🚀 30-second quickstart

# 1. Tools
brew install uv docker gh
open -a Docker

# 2. Project
git clone https://github.com/ZaltaClaw/company-brain.git ~/projects/company-brain
cd ~/projects/company-brain
uv sync
cp .env.example .env          # fill in values — see CLAUDE.md

# 3. Boot the brain
./scripts/init-db.sh          # postgres+pgvector on :5433
uv run company-brain status

# 4. Ingest + query
uv run company-brain ingest all
uv run company-brain search "what did I work on with Michal last week"

Full setup walkthrough lives in CLAUDE.md — written so Claude Code (or any agent) can finish the install on a fresh org Mac without human help.

---

🏗 Architecture

  ┌──────────────┐    ┌──────────────┐    ┌──────────────┐
  │  Fabric IQ   │    │   Work IQ    │    │  Foundry IQ  │
  │ (warehouse,  │    │ (mail, teams,│    │ (AI project  │
  │  lakehouses) │    │  sharepoint) │    │  knowledge)  │
  └──────┬───────┘    └──────┬───────┘    └──────┬───────┘
         │                   │                   │
   fabric_iq/            work_iq/           foundry_iq/
    ingest.py             ingest.py           ingest.py
         │                   │                   │
         └───────────────────┼───────────────────┘
                             ▼
                    ┌────────────────┐
                    │  extraction.py │  ← OpenAI structured output
                    │  embeddings.py │  ← text-embedding-3-small
                    └────────┬───────┘
                             ▼
                  ┌────────────────────┐
                  │    MemoryStore     │  packages/core
                  │  Postgres + pgvec  │  cosine search
                  │  container_tag RBAC│
                  └─────────┬──────────┘
                            │
                  ┌─────────▼──────────┐
                  │      FastAPI       │  apps/api
                  │ /v1/add  /v1/search│
                  │      /v1/profile   │
                  └─────────┬──────────┘
                            │
            ┌───────────────┼───────────────┐
            ▼               ▼               ▼
        Hermes         Claude Code        Cursor
       (or anything that speaks HTTP)

Every memory row carries a container_tag like fabric:<workspaceId>:<lakehouseId>, work:<userId>, or foundry:<projectId> so you can scope, filter, and govern searches per source.

---

🔌 Connectors

Connector	Source	Pulls	Auth
fabric_iq	Microsoft Fabric REST + SQL endpoint	Workspaces, lakehouses, table schemas, sample rows, semantic-model FK graph (the "ontology")	MSAL service principal
work_iq	Microsoft Graph	Outlook mail, Teams channel messages, OneDrive recent docs, calendar events	MSAL service principal or device-code flow
foundry_iq	Azure AI Foundry	Projects, knowledge indexes, docs, chunks	DefaultAzureCredential (works with az login)

Each connector lives in connectors/<name>/ with its own README.md, client.py, ingest.py, and (where relevant) auth.py + webhooks.py. Real Microsoft endpoints, real auth flows, no mocks.

Add your own connector

The interface is one method:

from company_brain.store import MemoryStore

def ingest(store: MemoryStore):
    for record in pull_from_my_source():
        store.add(
            content=record.text,
            container_tag=f"my-source:{record.id}",
            metadata={"source_url": record.url, ...},
        )

Drop it under connectors/<name>/ingest.py, register the command in cli.py, and you're done.

---

📡 API

uv run uvicorn apps.api.main:app --port 8088

Method	Endpoint	Purpose
POST	/v1/add	Insert a document — auto-extracted into atomic facts, embedded, indexed
POST	/v1/search	Hybrid semantic search across the brain, filterable by container_tag prefix
POST	/v1/profile	Build a fresh "user profile" — static facts + dynamic recent activity
GET	/v1/documents	List + paginate stored documents
GET	/healthz	Liveness check

Example:

curl -X POST http://localhost:8088/v1/search \
  -H 'content-type: application/json' \
  -d '{"query":"sales pipeline architecture","container_tag":"fabric:","limit":5}'

---

🧰 What's in the box

company-brain/
├── packages/core/company_brain/   # MemoryStore + Postgres impl + extraction + embeddings + CLI
├── connectors/
│   ├── fabric_iq/                 # Fabric REST + ontology graph
│   ├── work_iq/                   # Microsoft Graph (mail, teams, drive, calendar)
│   └── foundry_iq/                # Azure AI Foundry knowledge indexes
├── apps/api/                      # FastAPI HTTP wrapper
├── vendor/supermemory/            # ⬅ Supermemory dashboard, MCP, browser ext (git submodule)
├── scripts/                       # init-db, ingest-all, status
├── docker-compose.yml             # pgvector/pgvector:pg16 on :5433
├── CLAUDE.md                      # full handoff doc for Claude Code / any agent
└── README.md

Supermemory dashboard (vendored)

The upstream Supermemory repo is pinned as a git submodule at vendor/supermemory/ so you get their dashboard, MCP server, browser extension, and Raycast extension out of the box.

By default the dashboard points at hosted api.supermemory.ai. Flip it at your local Company Brain API by editing vendor/supermemory/apps/web/.env:

NEXT_PUBLIC_BACKEND_URL=http://localhost:8088     # your local Company Brain

Then:

cd vendor/supermemory && bun install && bun run dev    # dashboard on :3000

Memory graph viz

The vendored apps/memory-graph-playground is wired to render Company Brain's local data — no Supermemory cloud account needed. Our FastAPI exposes a Supermemory-compatible POST /v3/documents/documents, and a runtime patch points the playground at http://localhost:8088 instead of api.supermemory.ai.

./scripts/run-playground.sh    # opens graph viz on :3004 wired to your local brain

The patch lives at vendor-patches/memory-graph-playground.patch and is applied at runtime by scripts/apply-vendor-patches.sh, so the submodule pointer never moves. Override the brain URL with COMPANY_BRAIN_URL=... if you run the API on a non-default host/port.

The stores/supermemory.py adapter lets you flip the engine the other way too — keep the Supermemory dashboard pointed at their hosted brain and use Company Brain just for governed local ingestion. Either direction works; pick whichever side of the boundary you want to own.

To clone with submodules:

git clone --recurse-submodules https://github.com/ZaltaClaw/company-brain.git
# or if you already cloned:
git submodule update --init --recursive

---

🛡 Design principles

1. Local-first. No data leaves your machine except for source APIs (Microsoft) and OpenAI embeddings + extraction. No third-party memory SaaS.
2. One interface, swappable backend. MemoryStore is an ABC. Postgres+pgvector is the default; adapters for hosted Supermemory, mem0, MemClaw ship as stubs you can fill in.
3. Container tags as RBAC. Every row is scoped (fabric:, work:, foundry:, or anything custom). Searches filter by prefix. No cross-tenant bleed.
4. Real auth, no shortcuts. MSAL for Microsoft, DefaultAzureCredential for Azure, device-code fallback for laptop dev. No fake tokens.
5. Structured extraction. Facts come out of OpenAI as a Pydantic schema — {fact, confidence, temporal_validity, type} — not free-text blobs.

---

🗺 Roadmap

• Postgres + pgvector store with cosine search
• Fabric IQ connector (workspaces → lakehouses → tables → ontology)
• Work IQ connector (mail, Teams, OneDrive, calendar)
• Foundry IQ connector (projects → knowledge indexes)
• FastAPI HTTP layer
• CLI (init-db, ingest, search, profile, status)
• Graph change-notification webhooks for real-time Work IQ sync
• Hermes Agent memory-provider plugin (auto-wire Company Brain as Hermes's memory tier)
• Claude Code skill (/brain slash command)
• Web dashboard for browsing + reviewing memories
• Multi-user container_tag enforcement at the API layer (JWT-scoped)
• MemClaw adapter (governed cross-fleet episodic tier)

---

❓ FAQ

Why not just use Supermemory hosted? You can — there's an adapter stub. But "Company Brain" specifically means your brain, on your infra, with your keys. Hosted memory SaaS is a single point of trust we wanted to avoid for org data.

Why Postgres + pgvector instead of a dedicated vector DB? Because every org already has Postgres ops figured out, pgvector is fast enough for sub-second search up to ~10M rows, and you get transactional integrity for the metadata table for free. Swap it later if you outgrow it — the MemoryStore interface is intentionally tiny.

Does this work without OpenAI? Today it requires OpenAI for embeddings + extraction. Local-model adapters (Ollama for extraction, bge-large for embeddings) are a planned drop-in — the embeddings/extraction modules are isolated.

Will this break when Microsoft renames Fabric / Work / Foundry IQ next quarter? Probably. Connectors are isolated for exactly this reason — when the rename happens, you patch one file.

Can my agent write to the brain too? Yes. POST /v1/add is symmetric with the connector path. Hermes / Claude Code can drop notes into the brain mid-conversation and recall them next session.

---

🤝 Contributing

PRs welcome — especially new connectors (Linear, Jira, Slack, Notion, Salesforce, custom internal tools). Open an issue first if it's a big change.

git clone https://github.com/ZaltaClaw/company-brain.git
cd company-brain && uv sync
uv run pytest

---

📜 License

MIT — see LICENSE.

---

Built for the era where every employee runs a fleet of agents. Give them one brain. Keep the boundaries.

⭐ If this saved you a week of plumbing, star the repo. ⭐