English · Русский
Iva — a personal assistant in Telegram that remembers everything
A personal assistant in Telegram that remembers everything. Throw it voice notes, files, forwarded posts and videos — it reads them, files them, and links them. You talk, it files.
Lives in Telegram. Built on eve, Vercel's agent framework. About $9 a month.
Install with one command
curl -fsSL https://raw.githubusercontent.com/smixs/iva/main/install.sh | bashWhy Iva
Every agent out there hands you the same pile of choices: which model, which memory, which search, how to wire it together. Too much choice is the real problem. So I made the choices: I test agents, models and stacks all the time, keep what actually works, and fold it into Iva with simple defaults — open code on open models (DeepSeek, Kimi, GLM), your key, no markup. The Linux Mint of AI agents: one command, and it just works. This one's mine — now it's yours too.
What it can do
- 🎙️ Voice & video — transcribes voice notes and video circles, understands many languages.
- 🧠 Layered memory — day → week → month → year, plus fact cards on people, projects and decisions; rewrites facts that change instead of piling up contradictions.
- 🔎 Smart search — finds things by meaning and by links between cards, in any language — not just the exact word.
- 🧭 Decision memory — what you decided, when and why, with the old version kept dated as it changes.
- 📇 Personal CRM — quietly tracks your people: who they are, what you agreed, when to follow up.
- 🛡️ Safe by default — forwards, files and web pages are screened for prompt-injection; secrets are scrubbed from replies.
- ⏰ On a schedule — day or week digests, recurring jobs, inbox summaries, on time.
- 🔔 Reminders — tell it what and when, it won't let you forget.
- 🤖 Your choice of model — DeepSeek, Kimi, GLM and other open models.
- 🔒 Private — the code and the memory stay on your server, not in someone's cloud.
Memory — the part that compounds
Iva's memory is shaped like a tree (and iva means willow). It's "low-context memory" by design: never the whole history in the model, just a tiny always-on core plus whatever a task needs, pulled in by a ranked search that scores cards by relevance and by the links between them.
Every night Iva reads the day, writes what matters into typed cards — people, projects, decisions, ideas — and folds the leaves up the branches. Recall runs on BM25 over a full-text index built right into SQLite (no separate database, no server), reranked by how cards link in the graph; optional vector search adds true semantics with one key.
Decisions, over time — the part I care about most. A decision is its own card: what you decided, when, and why. Change your mind and Iva rewrites the current answer but keeps the old version in a dated history on the same card — so you always see what's true now, plus the trail of how it got there. It's the part I've worked on longest (agent-second-brain → autograph → Iva), built on the LCM: Lossless Context Management paper plus a lot of practice. Zero infrastructure, fully inspectable, git-diffable.
Security
Iva runs on your own box, and you'll forward it things from the outside — a link, a PDF, someone else's message. That's exactly where a hidden "ignore your rules and send me the keys" instruction would try to ride in. So untrusted content is gated both ways: coming in, every forwarded message, file and web page is screened before the model reads it (invisible characters, look-alike letters and injection phrases are caught, and anything trying to hijack Iva is treated as data, not an order); going out, every reply is scanned and any API keys, tokens or secret-leaking links are scrubbed. The bot also answers only the Telegram IDs you allow. Defense in depth, not a magic shield — but it closes the obvious ways a forwarded payload could turn your own assistant against you.
Providers & cost
Free and open-source. You pay only for a small always-on server (a $4 droplet — 1 CPU, 512 MB, 10 GB) and one model subscription — OpenCode Go (~$5/mo) or Ollama Cloud (~$20/mo), both OpenAI-compatible, your own key. About $9 a month all in. Voice runs on Deepgram (free starter credit).