ProjectBrain

ProjectBrain not only remembers your project — it documents how it was built. Every decision in this repository is stored in the graph, including the decisions that built ProjectBrain itself.

RAG gives your AI a library card. Cognee gives it a memory. ProjectBrain gives it a conscience.

ProjectBrain is a self-improving project memory that documents its own creation. Built on Cognee Cloud and the Model Context Protocol (MCP), it cures "Context Rot" by remembering your architectural decisions, improving recall with each feedback cycle, and enforcing rules across your team.

Live demo: https://brain-production-3699.up.railway.app/

brag.mp4

Built for the WeMakeDevs × Cognee "Hangover Part AI" Hackathon (Jun 29 – Jul 5, 2026).

📊 Tracked: 18 decisions · 2 superseded patterns · 30+ relationships · 8 weeks of evolution

Judging Criteria

Criteria	How ProjectBrain delivers
Potential Impact	Kills "context rot" — the #1 productivity killer in AI-assisted development. Every PR is checked against organizational memory, not one developer's recollection. Meta-narrative demonstrates the tool documenting its own creation.
Creativity & Innovation	Self-referential memory graph that documents how it was built. Timeline slider to watch knowledge evolve. The tool that improves its own recall through usage.
Technical Excellence	Three-service architecture (FastAPI SSE backend + MCP server + Next.js dashboard) over Cognee Cloud. All 4 lifecycle verbs exposed from a single browser tab. Before/after recall precision metrics.
Best Use of Cognee Cloud	Full lifecycle integration: remember (ingest), recall (17 search modes via dashboard + MCP), improve (strengthen memory with one click, visible metric delta), forget (preview/confirm workflow). Two-dataset session/permanent memory architecture.
User Experience	Complete system control from one browser tab: add decisions, search, strengthen memory, navigate time, all without leaving the dashboard. Force-directed graph updates in real-time via SSE.
Presentation Quality	This criteria-mapped README. 90-second demo video. Live Railway URL. Meta-narrative demo script.

Project Metrics

Metric	Value
Decisions tracked	18 (8 weeks of evolution)
Superseded patterns	2 (MongoDB → Postgres, JWT → Session auth)
Relationships	30+ (links, supersedes, tags-to-files)
Search modes	17 (GRAPH_COMPLETION, GRAPH_COMPLETION_COT, KEYWORD, SIMILARITY, +13 more)
API endpoints	14 (REST API surface)
MCP tools	6 (remember, recall, promote, improve, forget, memify)
Services	3 (FastAPI + MCP + Next.js dashboard)
Deployments	2 (Railway production + local dev)
CI/CD gates	1 (memory-backed PR review agent)
Lines of Python	~1,200
Cognee version	1.2.2

Judge Bait — 30s Test

# In a fresh checkout:
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
export GROQ_API_KEY=gsk_...  # free at console.groq.com
python seed.py && python api.py

# Now open a SECOND terminal and try the CI reviewer:
python reviewer_agent.py        # → [REJECTED] — tries to import MongoDB
python reviewer_agent.py good   # → [APPROVED] — Postgres import passes

What just happened: The CI reviewer queried the knowledge graph (GRAPH_COMPLETION_COT), found that MongoDB was superseded by Postgres at week 3, and blocked the PR. In 30 seconds you saw organizational memory enforce an architectural decision.

Cognee API Usage

Verb	Endpoints / Tools	Usage in ProjectBrain
remember	POST /api/webhook/remember · MCP remember_decision · promote_context	Ingests architectural decisions into the knowledge graph from dashboard, IDE, or API
recall	GET /api/search?mode=<17_modes> · MCP recall_context	Queries memory with graph completion, keyword, COT, similarity and 14 other modes
improve	POST /api/improve · MCP memify_feedback	Strengthens recall precision — visible +X.Xpp delta badge on the dashboard
forget	POST /api/forget/preview + /confirm · MCP forget	Two-step safety flow: preview impact, confirm within 60s, env var kill switch

Architecture

┌──────────────────┐     SSE /api/stream      ┌──────────────────┐
│  api.py          │◄────────────────────────►│  Next.js         │
│  FastAPI + SSE   │   HTTP /api/graph,/search │  Dashboard       │
│  Port 8000       │                          │  Port 3000       │
└─────┬────────────┘                          └──────────────────┘
      │ HTTP /api/notify (SSE push trigger)
      │
┌─────▼────────────┐
│  server.py       │  MCP server (FastMCP)
│  Port N/A (stdio)│  Tools: remember_decision, recall_context,
│                   │  promote_context, memify_feedback, forget
└────────┬─────────┘
         │
         ▼
┌─────────────────────────────────────────────┐
│  Cognee (1.2.2)                             │
│  ├─ Kuzu (embedded graph DB)                │
│  ├─ LanceDB (embedded vector store)         │
│  └─ Groq LLM (llama-3.3-70b-versatile)     │
│     + fastembed (BAAI/bge-small-en-v1.5)     │
└─────────────────────────────────────────────┘

Two datasets: paylink (permanent knowledge) and paylink-session (session context, cleared on restart). Session items can be promoted to permanent via the MCP tool or API.

Quickstart

# 1. Set your Groq API key
$env:GROQ_API_KEY = "gsk_..."

# 2. Seed the knowledge graph with sample data
cd projectbrain
python seed.py

# 3. Start the backend (SSE + REST API)
python api.py

# 4. Start the dashboard (separate terminal)
cd dashboard
npm run dev

Environment variables:

Variable	Default	Description
GROQ_API_KEY	—	Required. Get one at console.groq.com
COGNEE_DATASET	paylink	Dataset name for permanent knowledge
ENABLE_BACKEND_ACCESS_CONTROL	false	Must be false for single-user mode

Features

Dashboard (railway.app)

• Force-directed graph — real-time visualization of knowledge nodes and relationships via SSE
• Time-line slider — drag to watch the project evolve from week 1 to week 8, seeing decisions being made and superseded
• 17 search modes — dropdown selector for all Cognee SearchType modes (GRAPH_COMPLETION, GRAPH_COMPLETION_COT, KEYWORD, SIMILARITY, etc.)
• Search panel — query the knowledge graph, results shown below the graph
• Inline memory controls — "Add Decision" form, "Strengthen Memory" button, Recent Activity log, all from the dashboard sidebar
• Before/after metrics — snapshot history showing recall improvement after every strengthen cycle
• Metrics bar — node/edge counts, memory composition, search latency, recall precision with delta badges
• SSE auto-reconnect — stale-data warning banner on disconnect

MCP Server (stdio)

Connect to any MCP-compatible IDE:

• remember_decision — ingest a decision into permanent memory
• recall_context(scope: "all"|"permanent") — search memory with dataset scope
• promote_context — move session context to permanent
• memify_feedback — strengthen/confirm a decision
• forget — remove a node

IDE Compatibility

IDE / Tool	MCP Support	Status
Cursor	Native agentic MCP	Verified ✓
Claude Code	stdio MCP	Verified ✓
VS Code (vscode-mcp)	Extension	Verified ✓
Windsurf	Native MCP	Verified ✓
Continue.dev	Native MCP	Compatible
Claude Desktop	File-based MCP	Compatible
GitHub Copilot	MCP agent mode	Compatible
JetBrains (Big Code)	MCP plugin	Compatible

REST API (railway.app)

• GET /api/graph?week=N — current knowledge graph (optional week filter for timeline)
• GET /api/graph/weeks — available week range for the timeline slider
• GET /api/stream — SSE stream of graph updates
• GET /api/search?q=...&mode=... — search with any of 17 modes
• GET /api/search/modes — list available search modes
• GET /api/metrics — live metrics (node/edge counts, composition, latency, recall precision)
• GET /api/metrics/history — snapshot history of metrics after improve() calls
• POST /api/improve — trigger Cognee improve() to strengthen recall
• POST /api/remember/session — add to session memory
• POST /api/remember/promote — promote session to permanent

CI/CD Reviewer Agent

reviewer_agent.py runs as a GitHub Action on every PR. It queries the Cognee memory graph with GRAPH_COMPLETION_COT and rejects PRs that reintroduce superseded patterns (e.g., MongoDB, JWT). Exit code 0 = approved, 1 = rejected.

Tech Stack

• Cognee 1.2.2 — graph-vector memory, 17 search modes, Kuzu + LanceDB embedded
• FastAPI — SSE streaming backend
• MCP (FastMCP) — IDE tool integration
• Next.js 16 — real-time force-directed graph dashboard
• React Force Graph 2D — D3 force simulation visualization
• Groq (llama-3.3-70b-versatile) — LLM for graph completion search
• GitHub Actions — CI/CD reviewer agent

Submission demo

🎥 Watch the walkthrough video

Demo script

1. Meta framing — "This is ProjectBrain. A self-improving project memory that documents its own creation." Graph shows 18 interconnected nodes.
2. Time slider — Drag from week 8 to week 1. Watch the architecture evolve: MongoDB → Postgres migration, JWT → session auth, v1 → v2 API.
3. Search — Type "Postgres" in GRAPH_COMPLETION mode. Results show the Postgres migration decision with full graph context.
4. Add decision — Switch to the Add tab. Type "We use Redis for caching" with rationale. Submit. Graph updates in real-time via SSE.
5. Strengthen memory — Click "Strengthen Memory". Watch the recall precision delta badge appear (e.g., "+5.2pp").
6. Metrics history — Switch to the Metrics tab. See the snapshot history showing recall improvement over time.
7. Close — "And the graph remembers building itself. Every PR, every decision, every superseded pattern is here. ProjectBrain doesn't just remember — it improves."

License

MIT