cognee-companybrain

Unify fragmented company context — Slack threads and Granola meeting notes — into a single Cognee knowledge graph. From there, run agents on top: support assistants that recognize repeat customer questions, expert finders that route a question to the right person, contradiction detectors that surface where opinions diverge.

🚀 Start here

New to this? Two hands-on tutorials:

1. Learn Cognee — tutorial/intro_to_cognee.ipynb. Builds a tiny company brain from sample Slack + Granola data and walks through the core loop: remember → recall → node sets → typed schema → feedback. Runs in-process (no server). Quick start:

   bash tutorial/setup.sh            # creates .venv, installs deps
   # add LLM_API_KEY to .env
   source .venv/bin/activate && jupyter lab tutorial/intro_to_cognee.ipynb

   No code? Just open tutorial/graph*.html in a browser — see tutorial/GRAPHS.md.

2. Build the live Slack bot — see BOT.md.

The rest of this README describes the production ingestion pipeline.

---

Why

Knowledge in modern companies lives in too many places. Slack threads decay. Meeting notes sit unread. Decisions get re-litigated because nobody can find the prior conversation. One graph, fed continuously from the places people actually talk, fixes the retrieval problem.

What it does (v0)

1. Ingests Slack threads and Granola notes from configured sources
2. Normalizes them into a single transcript-shaped document with speaker, timestamp, and channel/meeting tags
3. Writes each document to Cognee via remember() with node_set tags carrying source/channel/project/speaker facets
4. Surfaces the graph through Cognee's built-in UI; add agents on top via recall()

Design

sources/slack.py     → fetch_threads(channel, since)     → ThreadData
sources/granola.py   → fetch_notes(since)                → NoteData
normalize.py         → to_doc(raw)                       → Doc
cognee_client.py     → write(doc)                        → remember(...)
scripts/run_ingest.py                                    → loop over sources

Every doc, regardless of source, looks like a transcript:

# <title>
[alice@acme.com, 2026-05-28T10:14] We have an auth bug on login.
[bob@acme.com,   2026-05-28T10:16] Same issue we hit last month?
[alice@acme.com, 2026-05-28T10:18] Yes — I'll patch it today.

Tags carry the structure: source:slack, slack:#support, doc:T123, speaker:alice@acme.com, optionally project:onboarding.

Cognee's default extraction reads the transcript with full context and produces Person, Topic, Decision, Question nodes with proper attribution. Custom DataPoint types can be layered on later when extraction quality plateaus.

Quickstart

# 1. Install
uv venv && source .venv/bin/activate
uv pip install -e ".[dev]"

# 2. Configure
cp .env.example .env
# Fill in SLACK_BOT_TOKEN, GRANOLA_API_TOKEN, COGNEE_API_KEY

# 3. Run ingestion
python scripts/run_ingest.py

Roadmap

• v0 — Slack + Granola ingest into a single Cognee dataset, default extraction
• v0.1 — Person/Project DataPoints synced from a source-of-truth (HR, Slack profile API)
• v0.2 — Demo agents: support agent, expert finder
• v0.3 — Recency-aware queries and contradiction detection
• v1.0 — Continuous ingest (webhook-driven), per-tenant access control

License

Apache 2.0 — see LICENSE. (Same license as cognee.)