Second Brain Summarizer

AI agent that reads daily message dumps from Google Drive, then summarizes, categorizes, and organizes them into a living knowledge base using the PARA method.

Built with LangChain + LangGraph for agentic reasoning, OpenRouter for model-agnostic LLM access, and Google Drive service account for file I/O.

How It Works

1. Fetches today's dump file (YYYY-MM-DD.md) from an input Google Drive folder
2. Parses messages delimited by <!-- msg_id: {id} --> markers
3. An AI agent reads the existing knowledge base structure, then autonomously:
   • Classifies each message into a PARA section (to-do, projects, areas, resources, archives)
   • Routes messages to the appropriate topic folder within that section
   • Merges new content with existing notes (read-before-write)
   • Updates each section's directory.md after every write
   • Updates the root directory.md when the top-level structure changes

Output Structure

The knowledge base uses a 3-level PARA hierarchy:

Output Drive Folder/
├── directory.md              # Root index of all 5 sections
├── to-do/
│   ├── directory.md          # Lists topic folders in this section
│   └── shopping/
│       └── groceries.md
├── projects/
│   ├── directory.md
│   └── dashboard-redesign/
│       └── notes.md
├── areas/
│   ├── directory.md
│   └── health/
│       └── running-log.md
├── resources/
│   ├── directory.md
│   └── books/
│       └── data-intensive-applications.md
└── archives/
    └── directory.md

Prerequisites

• Python 3.11+
• OpenRouter API key
• Google Cloud service account with Drive API enabled
• Two Google Drive folders (input and output) shared with the service account email

Quick Start

# Clone and set up
git clone <repo-url>
cd second_brain_summarizer
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

# Configure
cp .env.example .env
# Edit .env with your credentials

# Run
python -m second_brain.main                              # Process today's dump
python -m second_brain.main --date 2026-03-01            # Process a specific date
python -m second_brain.main --schedule                   # Run on cron schedule
python -m second_brain.main --prompt "list all projects" # Ad-hoc query

Configuration

Secrets go in .env:

Variable	Description
OPENROUTER_API_KEY	Your OpenRouter API key
GOOGLE_SERVICE_REFRESH_TOKEN	Path to refresh token JSON file
INPUT_DRIVE_FOLDER_ID	Google Drive folder ID containing dump files
OUTPUT_DRIVE_FOLDER_ID	Google Drive folder ID for the knowledge base

Non-secrets go in config.yaml:

llm:
  model: ""  # Any OpenRouter model
  temperature: 0.3
  max_tokens: 4096

seed_categories:
  - name: "work"
    description: "Work-related tasks, meetings, projects"
  - name: "personal"
    description: "Personal notes, reminders, ideas"

schedule:
  cron: "0 8 * * *"  # Daily at 8 AM

Dump File Format

Place markdown files named YYYY-MM-DD.md in the input folder. Messages are delimited by HTML comments:

<!-- msg_id: msg-001 -->
Had a productive meeting with the design team today.

<!-- msg_id: msg-002 -->
Finished reading chapter 5 on replication.

Docker

docker build -t second-brain .
docker run --env-file .env -v ./token.json:/app/token.json second-brain
docker run --env-file .env -v ./token.json:/app/token.json second-brain --schedule

Architecture

src/second_brain/
├── main.py          # Pipeline orchestrator + CLI
├── core/            # Config (Pydantic Settings) and shared data models
├── agent/           # LLM factory, system prompt, LangGraph ReAct agent
├── utils/           # Dump file parser
├── services/        # Google Drive API client (raw wrapper)
└── tools/           # LangChain @tool definitions (agent-facing)

The services/ + tools/ two-layer pattern is extensible: future integrations (Telegram, Google Calendar, etc.) add a new client in services/ and corresponding tools in tools/. The agent decides when to invoke them.