agent-second-brain

[!IMPORTANT] 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): «Созвон с Алишером — они согласны на пилот,
      но просят сдвинуть старт на июль. Надо обновить КП и напомнить мне
      в пятницу отправить договор.»

Bot:  💾 Сохранил: CRM-карточка Алишера обновлена (пилот, старт июль),
      связал с [[proekt-pilot]]. Напоминание поставлено: пятница 10:00 —
      отправить договор.

      — Friday, 10:00 —

Bot:  🔔 Напоминание: отправить договор Алишеру. Контекст: пилот, старт июль,
      КП обновлено во вторник.

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

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.

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

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:

c