Skip to content

smixs/agent-second-brain

Repository files navigation

Agent Second Brain

License: MIT Stars Forks

Agent Second Brain

Send a voice note to Telegram. Get an organized knowledge base, completed tasks, and a daily report back. That's it. That's the whole idea.

🇷🇺 Читать на русском


The problem

Every productivity system dies the same way. You set it up on a Sunday, use it for two weeks, then slowly stop because the overhead of maintaining it is more work than the work itself.

You have notes everywhere. Voice memos you never re-listen to. Ideas that disappear into chat history. Tasks you forget to write down. And even when you do capture something, it sits in a folder you'll never open again.

The real issue: organizing takes more effort than thinking. So the thinking never gets organized.

What this actually does

You talk to a Telegram bot. Voice, text, photos, forwarded messages - whatever is natural. You don't think about categories, tags, or where things go.

The agent handles everything else:

  • Transcribes your voice notes (Deepgram, takes seconds)
  • Classifies each entry - task, idea, client note, goal update, random thought
  • Creates tasks in Todoist with the right priority and due date
  • Saves everything to an Obsidian vault with proper links and tags
  • Sends you a daily report at 9pm - what happened, what got done, what's still hanging
  • Remembers what matters, forgets what doesn't - memory fades over time like a real brain

The bot runs 24/7 on a $5 VPS. You don't maintain it. You just talk to it.

Talk to it like a person

This isn't a button-pressing bot. You have a conversation.

You: what did I write about the marketing project last week?

Bot: finds and shows relevant entries

You: turn the second idea into a task for Monday

Bot: creates the task in Todoist

You: actually make it high priority and add a subtask for the presentation

Bot: updates the task

It has access to your entire vault, all your goals, and your Todoist. Ask it anything about your own notes, and it'll find the answer.

Memory that works like memory

Most AI systems either remember everything forever (drowning you in noise) or forget everything between sessions.

Agent Second Brain uses the Ebbinghaus forgetting curve - the same model that describes how human memory works. Every piece of information starts strong and gradually fades unless you access it again.

Five tiers, from always-on to nearly forgotten:

Tier What happens
Core Always in context. Your current projects, active clients, key goals.
Active Checked regularly. Recent ideas, ongoing conversations.
Warm Found when you search. Last month's notes, past decisions.
Cold Only surfaces in deep searches. Old projects, archived plans.
Archive Almost gone - but sometimes randomly recalled for creative connections.

The archive tier is the interesting one. Occasionally, the agent pulls a random old memory and connects it to something current. Sometimes it's noise. Sometimes it's the best idea you forgot you had.

Vault health - your notes maintain themselves

Over time, note systems rot. Links break. Files become orphans. Tags diverge. You end up with a graveyard of markdown files that nobody, including you, can navigate.

The vault-health system runs automatically:

  • Scores your vault on a 100-point scale
  • Finds orphan notes (no links in or out) and suggests connections
  • Repairs broken wiki-links
  • Generates Maps of Content (MOCs) for each domain
  • Flags files missing descriptions

You don't run maintenance. The agent does.

What you send, what happens

You send Agent does
Voice note about a client call Transcribes, creates CRM card, adds follow-up task
Quick text: "idea for the Q2 campaign" Saves to ideas folder, links to related notes
Forwarded article from a chat Saves with source, extracts key points
Photo of a whiteboard Saves with AI-generated description
"Process" button Runs the full pipeline right now
"What are my priorities this week?" Reads your goals and Todoist, gives you a straight answer

How it works (for the curious)

The daily processing runs in three phases:

  1. Capture - reads today's entries, classifies each one (task? idea? CRM update? goal progress?)
  2. Execute - creates Todoist tasks, writes vault files, updates cards
  3. Reflect - generates a summary report, updates long-term memory, sends it to Telegram

Each phase produces a clean JSON that the next phase picks up. If something breaks, you can see exactly where and why.

Telegram → Deepgram → Claude Code → Todoist + Obsidian vault → Telegram report

What it costs

Service Cost
Claude Pro $20/mo
VPS (any cheap one works) ~$5/mo
Deepgram Free tier ($200 credit)
Todoist Free plan works
Total ~$25/mo

$25/month for a personal assistant that organizes your life, never sleeps, and gets better the more you use it.

Quick start

1. Fork this repo

Click Fork at the top of this page. Make it private - it will contain your personal data.

2. Clone it

git clone https://github.com/YOUR_USERNAME/agent-second-brain.git
cd agent-second-brain

3. Fill in your info

Open these files and replace the placeholders:

  • vault/goals/ - your vision, yearly goals, monthly priorities, weekly focus
  • vault/.claude/skills/dbrain-processor/references/about.md - tell the agent about yourself
  • vault/.claude/skills/dbrain-processor/references/classification.md - how you want entries sorted

4. Get four API keys

What Where Time
Telegram Bot Token @BotFather 2 min
Your Telegram ID @userinfobot 30 sec
Deepgram API Key console.deepgram.com 3 min
Todoist API Token Todoist → Settings → Integrations → Developer 1 min

5. Deploy

Follow the VPS setup guide, or run:

ssh root@YOUR_SERVER_IP
curl -fsSL https://raw.githubusercontent.com/YOUR_USERNAME/agent-second-brain/main/bootstrap.sh | bash

That's it. The bot starts, you send it a message, and the system is alive.


Vault structure

vault/
├── daily/              # Your daily entries (voice, text, photos)
├── goals/              # Vision → yearly → monthly → weekly
├── business/
│   ├── crm/            # Client cards
│   └── network/        # Professional contacts
├── projects/           # Client work, leads, pipeline
├── thoughts/
│   ├── ideas/          # Ideas and brainstorms
│   ├── learnings/      # Lessons learned
│   └── reflections/    # Personal reflections
├── MOC/                # Maps of Content (auto-generated)
└── MEMORY.md           # Agent's long-term memory

Skills

The agent has five built-in skills:

Skill What it does
dbrain-processor Classifies entries, creates tasks, saves notes
agent-memory Ebbinghaus decay engine - remembers, forgets, recalls
vault-health Scores vault health, fixes links, generates MOCs
graph-builder Maps relationships between notes, finds clusters
todoist-ai Manages tasks, projects, priorities

Want just the memory engine? See agent-memory-skill - works standalone, no dependencies.

Configuration

File What it controls
.env API tokens (copy from .env.example)
.memory-config.json How fast memories decay, tier boundaries
mcp-config.json External tool connections
vault/.claude/CLAUDE.md Agent personality and rules

All secrets stay in .env, which is gitignored. Don't commit tokens.


Built by

Serge Shima - 20 years in marketing (BBDO, Publicis), now running an AI creative agency in Central Asia and teaching businesses how to work with AI at aimasters.me.

This system runs my actual life. 1,100+ vault cards, 5 AI agents, daily reports. It started as a weekend project and became infrastructure.

License

MIT - do whatever you want with it.

About

Send voice notes to Telegram → get organized knowledge base, tasks in Todoist, and daily reports. Persistent memory with Ebbinghaus decay, vault health scoring, knowledge graph. Runs on Claude Code + OpenClaw. 5/mo.

Topics

Resources

License

Stars

Watchers

Forks

Packages