Agent Second Brain
An always-on second brain you talk to.
Voice note in Telegram → typed, linked knowledge in your Obsidian vault.
Runs 24/7 on a $5 VPS and the Claude subscription you already pay for — zero per-token API bills.
🇷🇺 Русский •
Полная инструкция •
VPS guide •
autograph engine
Since June 15, 2026, headless claude -p runs bill against a separate paid Agent SDK credit. v3.0 sidesteps that completely: it drives one long-lived interactive Claude Code session — the same thing you run in a terminal, used exactly the way the subscription is meant to be used. No headless calls in the hot path (enforced by a CI guard), no per-request billing, no surprise invoice.
Start here
Why I built this
Every productivity system dies the same way. You set it up on a Sunday, use it for two weeks, then slowly stop — because maintaining the system is more work than the work itself. Voice memos you never re-listen to. Ideas that drown in chat history. A vault of markdown files that nobody, including you, can navigate a month later.
The fix isn’t another app. It’s removing the organizing step entirely: you talk, the agent files. And it has to be yours — I wasn’t going to pipe my private notes, clients, and goals through somebody’s SaaS. Everything here runs on your own server, lands in your own Obsidian vault as plain markdown, and is small enough to actually read before you trust it with your life.
What talking to it looks like
You (voice, 40s, while walking): "Call with Alisher — they're in for the
pilot, but want to push the start to July. I need to update the
proposal, and remind me Friday to send the contract."
Bot: 💾 Saved: Alisher's CRM card updated (pilot, July start), linked
to [[pilot-project]]. Reminder set: Friday 10:00 — send the contract.
— Friday, 10:00 —
Bot: 🔔 Reminder: send Alisher the contract. Context: pilot, July start,
proposal updated on Tuesday.
You: what did I write about the marketing project last week?
Bot: *finds the entries, quotes them, links the cards*
You: turn the second idea into a project note with next steps
Bot: *creates the note, links it to the client and this week's goals*
You: *forwards a post, drops a photo of a whiteboard, sends a PDF*
Bot: *reads them itself — files the takeaways into the graph, answers what it saved*
No commands to memorize, no categories to pick, no app to open. Telegram is the whole interface.
Philosophy
Voice-first. Capture has to be cheaper than forgetting, or the system dies. A voice note costs five seconds.
The vault is the source of truth. Everything lives as plain markdown in your Obsidian vault on your server. Delete the agent tomorrow — you keep everything, readable forever. No lock-in, no export button needed.
Memory that forgets, like yours. Storage is not memory. Knowledge decays on the Ebbinghaus curve, fades through five tiers, and resurfaces when it matters — so the graph stays sharp instead of becoming a landfill.
Interactive session, by the rules. One persistent Claude Code session in tmux, driven the way a human drives it. No headless claude -p anywhere in the hot path — a CI guard fails the build if anyone tries.
Small enough to read. One Python process, a handful of modules, 220+ tests. You can audit the thing that reads your private notes in an evening.
What it does
| 🎙 Total capture | Voice (Deepgram, seconds), text, photos, documents, videos, forwarded posts, whole albums — the agent reads files itself and files the takeaways. Nothing you send is ever silently dropped. |
| 🧠 Knowledge graph memory | Powered by autograph: typed cards, wiki-links, Ebbinghaus decay across five tiers, automatic MOC indexes, health scoring, link repair, dedup. The most thorough open-source memory layer you can drop on a vault. |
| ⏰ Self-managed routines | "Remind me Friday at 3pm", "every weekday at 18:30 check my inbox folder" — the agent schedules its own cron jobs from plain language. One-shots, intervals, full cron expressions. No external task manager. |
| 🌙 Nightly processing | At 21:00 your time it classifies the day's entries, writes vault cards, updates goals and long-term memory, rebuilds the graph — and sends you a daily report. |
| 🔌 Claude Code, but for PKM | It IS Claude Code under the hood — so drop any MCP server into mcp-config.json, add any skill into vault/.claude/skills/, and your second brain grows new abilities. Like Claude Code through Telegram — for knowledge, not code. |
| 🩺 Self-healing | A watchdog recovers a wedged session, a daily doctor sends a 🟢/🔴 canary report, broken jobs disable themselves and tell you. It just stays up. |
The memory engine: autograph
The part that makes this a brain rather than a logger is autograph — a typed memory layer for always-on agents, shipped here as a skill and also usable standalone on any Obsidian vault.
- Typed graph — every card carries a type (note, contact, project, CRM), a description for retrieval, tags, status; one
schema.json rules them all
- Ebbinghaus decay — strength
1 + ln(access_count): each touch slows forgetting; contacts fade in ~100 days, dailies in ~25
- Five tiers — core → active → warm → cold → archive; touching a card promotes it back up
- Random recall — occasionally an archived card resurfaces next to something current. Sometimes noise. Sometimes the best idea you forgot you had.
- Self-maintenance — orphan detection, broken-link repair, dedup into
.trash/, MOC generation, a 100-point health score. You never run vault chores again.
| Tier |
What happens |
| Core |
Always in context: current projects, active clients, key goals |
| Active |
Checked regularly: recent ideas, ongoing threads |
| Warm |
Found when you search |
| Cold |
Surfaces only in deep searches |
| Archive |
Almost gone — but randomly recalled for creative collisions |
How it works
Telegram ──▶ bot (aiogram) ──▶ persistent Claude Code session (tmux pane)
│ │
│ ├──▶ Obsidian vault (plain markdown)
│ └──▶ autograph: graph · decay · MOC
├──▶ cron ticker ──▶ second isolated session (reminders never block chat)
└──▶ watchdog + daily doctor (self-healing, 🟢/🔴 report)
The bot never spawns claude per message. It keeps one long-lived interactive Claude Code session alive in a tmux pane and types prompts into it — which is why a 24/7 agent runs on a flat Pro/Max subscription instead of metered API tokens. Scheduled jobs fire in a second, isolated session, so a reminder going off never interrupts your conversation. A cross-process lock serializes everything; a watchdog restarts whatever wedges.
Privacy is explicit: voice audio goes to Deepgram for transcription, text goes to Anthropic through your subscription, everything else stays on your server. The vault never leaves the machine.
What it costs
| Service |
Cost |
| Claude Pro |
$20/mo |
| VPS (any cheap one) |
~$5/mo |
| Deepgram |
free tier ($200 credit) |
| Total |
~$25/mo, flat |
A personal assistant that organizes your life, never sleeps, and gets better the more you use it — for the price of two coffees. And because the session is interactive, the price above stays the price: there is no token meter to run away from you.
Quick start
Three steps, ~15 minutes of work (full beginner walkthrough: 🇷🇺 setup-guide / 🇬🇧 vps-setup).
1. Fork this repo (make the fork private — it will contain your life), then fill in vault/goals/ and the persona in vault/.claude/skills/dbrain-processor/references/about.md.
2. Get two keys: a bot token from @BotFather, a free key from Deepgram, plus your Telegram ID from @userinfobot.
3. On a fresh Ubuntu VPS:
curl -fsSL https://raw.githubusercontent.com/YOUR_USERNAME/agent-second-brain/main/bootstrap.sh | bash
The installer interviews you for the tokens, walks you through Claude Code login (browser link), installs systemd services, and finishes with a health check. The bot messages you when it’s alive. That’s it — start talking.
What the installer actually does (inspect before you trust)
bootstrap.sh clones your fork and hands off to setup.sh, which only asks questions (tokens, timezone, git remote) and writes .env (chmod 600). All real work happens in idempotent upgrade.sh: uv + Python deps, tmux, Claude CLI, dbrain-* systemd user units, a dbrain CLI (status / logs / attach / doctor), permission hardening, and a first doctor run. Every script is short enough to read first.
Upgrading from v1 / v2
One idempotent command — it migrates old units, installs what’s missing, repairs permissions, and health-checks itself:
ssh your-server
cd agent-second-brain && git pull && bash upgrade.sh
If a migration ever goes sideways, run claude in the project directory and call the migrate-doctor skill — it diagnoses the install layout (v1/v2/v3), backs up, and repairs interactively.
Vault structure
vault/
├── daily/ # Your raw daily stream (voice, text, attachments)
├── goals/ # Vision → yearly → monthly → weekly
├── business/
│ ├── crm/ # Client cards
│ └── network/ # Professional contacts
├── projects/ # Active work, leads, pipeline
├── thoughts/
│ ├── ideas/ # Ideas and brainstorms
│ ├── learnings/ # Lessons learned
│ └── reflections/ # Personal reflections
├── MOC/ # Maps of Content (auto-generated)
└── MEMORY.md # The agent's long-term memory
Skills
| Skill |
What it does |
| dbrain-processor |
Classifies entries, writes vault cards, daily reports |
| autograph |
The memory engine: decay, graph health, MOCs, schema-as-code, dedup |
| cron |
The agent manages its own schedule from plain language |
| migrate-doctor |
Diagnoses and repairs broken upgrades from older versions |
Drop your own into vault/.claude/skills/ — the session picks them up like any Claude Code install.
Configuration
| File |
Controls |
.env |
Tokens, timezone, model, cron tuning (see .env.example) |
vault/.claude/CLAUDE.md |
The agent’s personality and rules |
mcp-config.json |
Optional: any MCP servers you want the brain to have |
All secrets stay in .env (gitignored, chmod 600). Set CLAUDE_MODEL=sonnet to reduce weekly-limit pressure on a 24/7 bot.
Set TZ to your own IANA timezone (e.g. TZ=Europe/Moscow) — it defaults to UTC, and reminders/schedules are interpreted in it. The cron skill’s examples use placeholders, not a fixed zone, so the brain schedules in your time.
dbrain restart does not recreate the brain. It restarts the bot process; the long-lived Claude Code tmux session deliberately survives (KillMode=process) — that’s what keeps you on the subscription. Settings read only at session creation (CLAUDE_MODEL, mcp-config.json, the system prompt) therefore take effect only after the brain is recreated. Kill both sessions — the main one and the cron brain (${brain}_cron, where reminders run with the same settings) — and the bot rebuilds them on the next message/job:
B="$(cat ~/.dbrain/brain.name)"; tmux kill-session -t "$B"; tmux kill-session -t "${B}_cron" 2>/dev/null
Bot-level .env changes (tokens, ALLOWED_USER_IDS, TZ) apply on a plain restart.
What it does NOT do
- It does not send your vault anywhere. Plain markdown, your disk, your git remote if you configure one.
- It does not use headless
claude -p — scripts/check-no-claude-p.sh runs in CI and fails the build if that ever changes.
- It does not require any API keys beyond Telegram + Deepgram. No OpenAI, no token meter.
- It does not lock you in. Delete everything tomorrow; your vault still opens in Obsidian.
Built by
Serge Shima — 20 years in marketing (BBDO, Publicis), now running an AI creative agency in Central Asia and teaching businesses to work with AI at aimasters.me.
This system runs my actual life: 1,100+ vault cards, daily reports, an agent I argue with about my own calendar. It started as a weekend project and became infrastructure.
License
MIT — do whatever you want with it.
评论