Your thoughts don't disappear here.
Most apps make you organize as you go. BrainDump lets you dump everything raw — text or voice — and the AI does the organizing. But what's different: it remembers. Every dump, every task, every connection you've ever made lives in context. So when you mention someone you met, or an idea you had, the AI already knows what's open, what's related, and what needs updating.
It's not a note app. It's not a task manager. It's a second brain that actually reads what you put in it.
Most note apps just store what you write. BrainDump is context-aware — when you dump a new thought, the AI already knows your existing pending tasks and recent dumps. So if you mention meeting someone who's a hiring manager, and you already have a task "Research security job opportunities", the AI links them — enriching the existing task rather than creating a duplicate.
Example: You dump "Met Jake at the networking event, he's a hiring manager at CyberX". The AI sees you already have a pending task "Research security job openings". Instead of creating a duplicate, it enriches the existing one: "Research security job openings — contact Jake from CyberX (met at networking event)".
The AI uses a three-tier retrieval system to find relevant tasks:
- Vector search (production) — cosine similarity via pgvector + BAAI/bge-small-en-v1.5 embeddings
- Full-text search — keyword matching via PostgreSQL
tsvector, always available as fallback - Recency — last N tasks by creation date, as a final safety net
| Feature | Details |
|---|---|
| Dump freely | Text or voice, no structure needed — just brain-dump and let the AI organize |
| AI task extraction | DeepSeek-V3 extracts tasks with priority, due dates, and context from your full history |
| Cross-dump memory | New dumps enrich existing tasks instead of creating duplicates |
| Voice input | Web Speech API (browser-native, zero cost) with HF Whisper fallback |
| Task management | Edit title, description, priority, status, due date, subtasks, notes, tags, schedule |
| Task filtering & sorting | Filter by priority/status, sort by newest / due date / priority |
| Bulk actions | Select multiple tasks, mark complete or delete in one click |
| Export | Download all tasks as CSV or JSON |
| Search | ⌘K global search across tasks and notes with keyword highlighting |
| Notes view | Every brain dump linked to the tasks it created — click to expand |
| Semantic retrieval | pgvector HNSW index surfaces relevant older tasks, not just the most recent |
| Rate limiting | 20 dumps/hour (DB-backed), 30 transcriptions/hour (in-memory) |
| Auth | Supabase Auth with session cookie refresh on every request |
This is a single full-stack Next.js app — there is no separate backend server.
| Layer | Technology |
|---|---|
| Frontend | React 19 + Next.js 16 App Router |
| UI | shadcn/ui (Radix UI) + Tailwind CSS v4 |
| Client data | SWR with optimistic updates |
| Backend | Next.js API Routes → Vercel serverless functions |
| AI (tasks) | deepseek-ai/DeepSeek-V3-0324 via Hugging Face router |
| AI (voice) | openai/whisper-large-v3 via Hugging Face (fallback) |
| Voice (primary) | Web Speech API — browser-native, free, no API call |
| Embeddings | BAAI/bge-small-en-v1.5 (384-dim) via Hugging Face Inference API |
| Database | Supabase (PostgreSQL + pgvector) |
| Auth | Supabase Auth, JWT, cookie refresh via middleware |
| Hosting | Vercel |
brain_dumps id uuid (PK) user_id uuid (FK → auth.users) content text created_at timestamptz tasks id uuid (PK) user_id uuid (FK → auth.users) brain_dump_id uuid (FK → brain_dumps) title text description text priority enum: low | medium | high status enum: pending | in_progress | completed due_date timestamptz subtasks jsonb notes text schedule_type text scheduled_date timestamptz tags text[] embedding vector(384) -- pgvector semantic search created_at timestamptz updated_at timestamptz -- kept accurate by DB trigger api_logs id uuid (PK) user_id uuid brain_dump_id uuid endpoint text model text content_length int tasks_extracted int enrichments_applied int duration_ms int success boolean error_message text created_at timestamptz
git clone https://github.com/ahmedthebutt/braindump.git
cd braindump
npm installCreate .env.local:
NEXT_PUBLIC_SUPABASE_URL=your_supabase_project_url NEXT_PUBLIC_SUPABASE_ANON_KEY=your_supabase_anon_key HUGGINGFACE_API_TOKEN=your_hf_token
- Supabase — create a free project at supabase.com. Get
SUPABASE_URLandSUPABASE_ANON_KEYfrom Project Settings → API. - Hugging Face — create a free account at huggingface.co and generate a token in Settings → Access Tokens. One token covers all three models (DeepSeek-V3 for task extraction, Whisper for voice, BGE for embeddings).
In the Supabase SQL editor, run these migrations in order:
supabase/migrations/001_task_enhancements.sql
supabase/migrations/002_updated_at_trigger_and_rls.sql
supabase/migrations/003_embeddings.sql
Migration 003 requires the pgvector extension (pre-installed on all Supabase projects). It adds the embedding column, HNSW index, and two RPCs (match_tasks for vector search, match_tasks_fts for keyword fallback).
npm run dev
Open http://localhost:3000.
User types or speaks a brain dump
↓
POST /api/extract-tasks
↓
Fetch context:
• Last 5 brain dumps (recent history)
• Up to 15 relevant pending tasks via three-tier retrieval:
Tier 1: pgvector cosine similarity (embed dump → find nearest tasks)
Tier 2: PostgreSQL full-text search (keyword overlap fallback)
Tier 3: Most recent by created_at (last-resort baseline)
↓
DeepSeek-V3 prompt:
STEP 1 — INVENTORY all items in the dump
STEP 2 — CLASSIFY each as: new task / enrichment / subtask / not a task
STEP 3 — OUTPUT: tasks[], enrichments[], subtask_additions[], summary
↓
Apply results:
• tasks[] → INSERT new rows into tasks table (with embedding)
• enrichments[] → UPDATE description of existing tasks
• subtask_additions[] → PUSH new subtask into existing task's jsonb array
• Log to api_logs
↓
Return to client → SWR mutate → task list re-renders
npm run dev # Start dev server npm run build # Production build npm run test # Run unit tests (Vitest, 24 tests) npm run test:story # Run 8-chapter AI story integration test (real HF + Supabase) npm run backfill:embeddings # Embed all existing tasks that predate migration 003
The project has two test layers:
Unit tests (npm run test) — 24 tests covering rate limiting, Zod response schemas, dedup logic, and utility functions. Fast, no external calls.
Story integration test (npm run test:story) — feeds an 8-chapter job-search story to the real AI pipeline and evaluates intelligence: does it enrich existing tasks? avoid duplicates? handle extreme dumps with 10 items? link information across dumps? Last run: 95% (36/38 checks passed).
app/
api/
extract-tasks/route.ts Core AI pipeline
transcribe/route.ts Voice → text via Whisper
tasks/[id]/route.ts PATCH + DELETE individual tasks
tasks/batch/route.ts Bulk complete + delete
tasks/export/route.ts CSV / JSON export
brain-dumps/[id]/route.ts Delete dump + cascade tasks
auth/ Login, sign-up, callback pages
dashboard/
page.tsx Server component: auth check + SSR data
dashboard-content.tsx Client shell: SWR, views, capture
page.tsx Landing page
components/
capture-zone.tsx Voice + text capture UI with mascot
task-list.tsx Task rows: status, priority, delete, open detail
task-detail-panel.tsx Slide-in edit panel
error-boundary.tsx React class error boundary
lib/
embeddings.ts HF embedding helpers (query + task embedding)
rate-limit.ts DB-backed + in-memory rate limiting
supabase/
server.ts Server-side Supabase client (cookie-based)
client.ts Browser Supabase singleton
supabase/migrations/ SQL migrations (run in Supabase SQL editor)
__tests__/ Vitest unit + story integration tests
scripts/ CLI utilities (backfill, screenshots)