v0.17.0 · open source · MIT

Ikeptlearningthingsandthenlosingthem.

So I built a system that remembers. Self-hosted AI memory, capture anything, search semantically, write in your voice.

Scroll
Primary interface

Talk to it on Telegram

Voice notes, links, text. Send anything, it figures out the rest.

10:35
E
EchOS
online
Message
21:15
E
EchOS
online
Message
echos · ~/knowledge
The core loop

Capture. Search. Write.

Three primitives. Extend everything else with plugins.

Capture

Send anything: voice notes, URLs, text, images, YouTube videos, tweets. EchOS extracts, summarizes, and tags it automatically.

Search

Hybrid search combining full-text (BM25) and semantic vectors with Reciprocal Rank Fusion. Find what you mean, not just what you typed.

Write

Generate blog posts, emails, and threads grounded in your notes, in your voice, trained on your writing examples.

Under the hood

How it works

An LLM agent with tools, not rigid command routing. Plain Markdown files you can open in Obsidian.

Architecture

inputTelegram · CLI · Web
agentClaude + tools
storageSQLite + LanceDB + Markdown
searchBM25 + semantic + RRF fusion
TypeScript
Node.js
SQLite
LanceDB
Claude
grammY
BullMQ
Redis
Fastify
Pino

Interfaces

stable
Telegram
Primary interface: voice, text, media
stable
CLI
Standalone terminal interface
experimental
Web
REST API + web UI

Plugins

Youtube
Transcript extraction
Article
Web article extraction
Content Creation
Blog posts, emails & threads
Image
Image storage & analysis
Digest
Daily knowledge digest
Twitter
Tweet thread extraction
Resurface
Rediscover forgotten notes
Journal
Periodic journaling
Pdf
PDF document extraction
Audio
Audio transcription & capture
Rss
RSS feed monitoring
Self-host in minutes

Get started

Your data stays yours. No subscriptions. No cloud dependency.

macOS, quickest path
brew install echos && echos-setup