Engram

Forward anything to Telegram. Get a tagged, linked, deduplicated Obsidian note back.

An engram is the physical trace a memory leaves in the brain — the durable scar left behind after an experience. Engram is a single-tenant Telegram bot that does the same thing for your chat stream: drop in a tweet, voice memo, PDF, YouTube link, or photo of a whiteboard, and Claude classifies it, summarises it, tags it, finds related notes already in your Obsidian vault, and writes a Markdown file with proper frontmatter and [[backlinks]]. The forgettable river of messages becomes durable, indexed memory.

Built as a personal second-brain pipeline; published in case it's useful to anyone else.

---

What it does

• Capture anything — text, URLs, photos, voice notes, PDFs, Word docs (.docx/.doc), plain-text/code files, forwarded posts, YouTube links. Media groups (multi-photo posts) are debounced and stitched into a single note.
• Read the contents, not just the message — Claude vision OCR for photos, OpenAI Whisper for voice, PyPDF/python-docx for documents, YouTube Transcript API for videos, plain HTTP fetcher for web pages.
• AI routing — Claude (Sonnet) picks a folder, writes a title, summarises the body, generates tags, and proposes up to 5 related notes from your existing vault. Falls back to an "Other" bucket and an inbox queue when confidence is low.
• Smart dedupe — incoming notes that match by URL, title, or semantic similarity (cosine on OpenAI embeddings) get appended to the existing note instead of creating a duplicate.
• Vault-grounded Q&A — /ask runs a hybrid keyword + embedding retrieval over your notes and answers with Claude, with multi-turn follow-ups via Telegram reply threads.
• Manual override — every capture shows an inline-keyboard folder picker; misroutes are one tap away. /redo, /edit, /undo, and /relink cover the rest.
• Single-tenant by design — an ALLOWED_USER_IDS allowlist gates every handler. Nobody else who finds your bot can use it.

How it works

Telegram message
     │
     ├─ photo?       → Claude vision OCR
     ├─ voice/audio? → OpenAI Whisper
     ├─ PDF / DOCX / DOC / text? → text extraction
     └─ URLs?        → page fetch · YouTube transcript
     ↓
Claude enrichment: title · summary · tags · folder · related notes · confidence
     ↓
Dedupe check (URL → title → semantic) → append to existing OR create new
     ↓
<vault>/<Category>/<Title>.md   with YAML frontmatter + [[backlinks]] + attachments/

Requirements

• Python 3.11+
• uv (or plain pip if you prefer)
• A Telegram bot token from @BotFather
• An Anthropic API key
• (optional) An OpenAI API key for voice transcription, semantic dedupe, semantic /search, and /relink
• An existing Obsidian vault (or any folder you want filled with .md files)

Setup

Install from PyPI

pip install engram-bot   # or: uv pip install engram-bot
# create .env in your working dir (see Configuration below), then:
engram

The PyPI distribution name is engram-bot (because engram is squatted on PyPI); the Python import name and the console script are both engram.

From source

git clone https://github.com/mishablank/Engram.git
cd Engram
uv sync
cp .env.example .env
# edit .env — see Configuration below
uv run python -m engram.bot

One-click deploy (Railway)

Click the Deploy on Railway button above. Railway will build the project, prompt you for the env vars below, and run uv run python -m engram.bot as a long-lived process.

The catch: your Obsidian vault is local, but Railway runs in the cloud. Two ways to make this work:

1. Recommended — attach a Railway volume mounted at e.g. /data, set BASE_DIR=/data, and use Obsidian Sync, Syncthing, or rclone to mirror that volume into your local Obsidian vault. The bot writes to the cloud copy; your desktop reads it via sync.
2. Quick test — point BASE_DIR at the container filesystem and treat it as ephemeral. Notes survive restarts but vanish if Railway redeploys without a volume. Fine for kicking the tyres, not for real use.

If you don't want any of that, just run it on your laptop or a home server. See "Running it as a daemon" below.

Configuration

All config is via environment variables (loaded from .env if present). See .env.example.

Variable	Required	Purpose
TELEGRAM_BOT_TOKEN	yes	Bot token from @BotFather
ANTHROPIC_API_KEY	yes	Used for enrichment, vision OCR, /ask
ALLOWED_USER_IDS	yes	Comma-separated Telegram numeric IDs. Only these users can talk to the bot. Find yours via @userinfobot.
BASE_DIR	yes	Absolute path to your Obsidian vault root
OPENAI_API_KEY	no	Enables Whisper voice transcription, semantic dedupe, /relink, and semantic ranking inside /search and /ask
CATEGORIES	no	Comma-separated folder names. Default: AI,Crypto,Startups/YC,Personal,Health,Reading,Other
LOG_FILE	no	Defaults to ~/.engram.log (rotated, 5 MB × 3 backups)

Commands

Command	Effect
/start	Show usage hint with the current category list
/search <query>	Hybrid (keyword + embedding) search across titles, tags, and bodies
/ask <question>	RAG over your vault. Reply to my answer to continue the thread (up to 6 turns)
/inbox	List notes flagged for review (low-confidence routing)
/review	Walk pending notes one at a time with move / mark-reviewed / delete buttons
/relink [folder]	Refresh related-note backlinks. No arg = last capture; with arg = entire folder
/redo	Reply with /redo to regenerate a capture using the higher-quality Opus model
/edit <text>	Replace the source of the last capture and re-enrich
/undo	Delete the last capture in this chat
/refresh	Rescan the vault index (also runs automatically every 10 minutes)

The plain message path: send a message → tap a folder button → done. Send a photo without a caption and the bot OCRs it first so it can route by content.

What a note looks like

---
title: "Notes on bitter-lesson scaling laws"
created: 2026-05-11T18:32:04
source: https://example.com/post
source_type: article
tags: [scaling-laws, rich-sutton, ai]
forwarded_from: "@somechannel"
forwarded_at: 2026-05-11T18:30:00
---

Sutton argues that the only methods that consistently win across decades
are those that scale with compute and data — search and learning — and that
hand-tuned domain knowledge tends to be a local optimum at best.

## Related
- [[The bitter lesson, revisited]]
- [[Compute overhang and capability surprises]]

![[attachments/2026-05-11_18-32-04-1.jpg]]

Running it as a daemon

The bot is a long-lived process. Some lightweight options:

• macOS (launchd): drop a ~/Library/LaunchAgents/com.you.engram.plist that runs uv run python -m engram.bot with KeepAlive=true.
• Linux (systemd user unit): a one-screen ~/.config/systemd/user/engram.service with ExecStart=… and Restart=on-failure, then systemctl --user enable --now engram.
• Quick and dirty: tmux new -d -s engram "uv run python -m engram.bot".

Logs go to LOG_FILE (default ~/.engram.log).

Development

uv sync
uv run pytest -v

15 test modules cover the bot handlers, vault indexing, embeddings, dedupe, link enrichment, vision/whisper/youtube/pdf adapters, and the inbox/review flow. pytest-asyncio is in auto mode. CI runs on push and PR against Python 3.11 / 3.12 / 3.13.

Roadmap

• Local-model support — swap Claude / OpenAI for Ollama or llama.cpp so the bot can run end-to-end without paid API keys. Embeddings first (cheapest win), then enrichment. Tracked in #1 — help welcome.

Security model

This bot is single-tenant on purpose. It does exactly one thing to keep you safe:

• Every handler checks update.effective_user.id against ALLOWED_USER_IDS before doing anything. Unauthorised users get a flat "Unauthorized." reply.

That's it. There is no per-user vault, no row-level auth, no rate limiting. Don't share your bot token. Don't add other users to ALLOWED_USER_IDS unless you want them writing into the same vault you do.

.env is gitignored. Treat your TELEGRAM_BOT_TOKEN and ANTHROPIC_API_KEY like passwords — if either leaks, rotate immediately (BotFather → /revoke, Anthropic console → revoke key).

License

MIT © 2026 Mikhail Blank