We taught the Drosophila fly to speak.

Previously, the fly could only do what it was told. But now the fly can also talk to people.

📜 License — read this before you fork

Engram is source-available, not OSI open-source software. The full terms are in LICENSE. What they mean for you:

• You may read and study the source code, and use, modify, and run the program for personal, non-commercial purposes.
• You may NOT fork, modify, publish, or redistribute the project, and may NOT use it commercially (selling it, offering it as a service, bundling it into a product) — without prior written permission from the author.
• Any permitted modification must retain the copyright notice and this license as-is.

In short: forking for a personal experiment is fine per the license; publishing your fork, redistributing it, or using it commercially requires asking first. To request permission, open an issue (github.com/nethoster/engram/issues).

What is Engram?

Engram is a fully local pipeline:

1. Connectome data — a subgraph of the male-cns:v1.0 dataset fetched from neuPrint (Janelia) via fetch_subgraph.py, stored as local parquet files. No live API calls at runtime.
2. Local LLM brain — Ollama running huihui_ai/qwen3-abliterated:0.6b-v2 (uncensored Qwen3, tool-calling capable).
3. Telegram bot — an aiogram 3 bot that routes questions to the LLM, which calls graph tools (neuron_info, shortest_path, hub_neurons, ...) and answers from real data.

The bot has an English / Russian interface — switch with the 🌐 Language button or /language.

Requirements

• Python 3.10+ (developed on 3.14)
• Ollama running locally with the model pulled:

  ollama pull huihui_ai/qwen3-abliterated:0.6b-v2

• A Telegram bot token from @BotFather (create your own: /newbot)
• A neuPrint API token from neuprint.janelia.org (free; only needed to fetch the connectome data)

Install

Option 1 — Universal (all OSes)

git clone https://github.com/nethoster/engram.git
cd engram
python3 -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install -r requirements-bot.txt

Option 2 — One command (Linux, macOS, Windows via Git Bash / WSL)

curl -fsSL https://raw.githubusercontent.com/nethoster/engram/main/install.sh | sh

The installer creates a virtualenv, installs dependencies, and walks you through entering your tokens (Telegram + neuPrint).

Setup

1. Telegram token (required)

Create a bot with @BotFather → /newbot → copy the token. Put it in a .env file in the project root:

TELEGRAM_BOT_TOKEN=123456:ABC-DEF...

2. neuPrint token (for fetching data)

Register at neuprint.janelia.org, generate an API token, then:

export NEUPRINT_APPLICATION_CREDENTIALS=<your-token>
python fetch_subgraph.py            # writes data/*.parquet

The fetched data (data/*.parquet) is local-only and not committed to git. Each user fetches it with their own token.

3. Run

./bot.sh start        # start polling
./bot.sh status       # check
./bot.sh stop         # stop
# or directly:
python -m bot.main

Usage

Open your bot in Telegram and:

• /start — main menu
• /language — switch EN / RU
• /reset — clear the LLM dialog
• Ask questions: who are the graph hubs?, path from 10009 to 10157, find type CT1

Project structure

engram/
├── bot/                    # Telegram bot (aiogram 3)
│   ├── brain/              # LLM client + system prompts (EN/RU)
│   │   ├── ollama.py       #   tool-calling loop
│   │   └── prompt.py       #   system prompts per language
│   ├── connectome/         # graph data + tools
│   │   ├── loader.py       #   parquet -> networkx DiGraph
│   │   ├── tools.py        #   6 tools exposed to the LLM
│   │   └── tool_specs.json #   Ollama tool schemas
│   ├── handlers/           # Telegram command & chat handlers
│   ├── ui/                 # formatting (Markdown -> Telegram HTML)
│   ├── i18n.py             # localization (EN/RU)
│   ├── state.py            # per-chat dialog + language
│   ├── config.py           # env / .env config
│   └── main.py             # entrypoint
├── data/                   # local subgraph (gitignored, fetch yourself)
├── fetch_subgraph.py       # neuPrint -> parquet (needs your token)
├── bot.sh                  # start / stop / status / restart
├── install.sh              # one-command installer
├── requirements-bot.txt    # Python dependencies
└── LICENSE                 # Source-Available License

Tech stack

• Python + aiogram 3 (Telegram)
• Ollama + huihui_ai/qwen3-abliterated:0.6b-v2 (local LLM, tool calling)
• networkx + pandas + pyarrow (graph & data)
• neuprint-python (data fetch only; not needed at runtime)

Notes

• Everything runs locally — no live neuPrint queries while the bot is running.
• The subgraph: 2000 neurons, 225085 directed edges, 1 weakly-connected component, density 0.056 (BFS from top-10 traced seeds by post).
• The default interface language for new chats is English; switch to Russian with 🌐 Language.