Skip to content

Telegram memory bot

The aiogram 3 demo is a real polling application with persistent SQLite memory. It automatically stores completed turns, recalls relevant old turns, keeps a small hot-history window, and exposes content-free diagnostics.

Install and run

Choose one model backend:

pip install "protoprompt[telegram,ollama]"
ollama pull llama3.1 nomic-embed-text
export TELEGRAM_BOT_TOKEN="..."
export PROTOPROMPT_PROVIDER="ollama"
python examples/telegram_memory_bot.py

pip install "protoprompt[telegram,openai]"
export TELEGRAM_BOT_TOKEN="..."
export OPENAI_API_KEY="..."
export PROTOPROMPT_PROVIDER="openai"
python examples/telegram_memory_bot.py

On PowerShell, use $env:NAME="value". PROTOPROMPT_DB selects the SQLite file and defaults to telegram_memory.db. Model names and endpoints can be changed with OPENAI_* or OLLAMA_* environment variables shown in the example source.

Do not reuse an existing database after changing to an embedding model with a different vector dimension. Use a new database or re-embed the stored memory.

Commands and privacy

  • /memory reports current-thread, all-thread, and hot-memory counts;
  • /why shows ids and similarity scores for the last recall, never text;
  • /forget explains the destructive action;
  • /forget confirm deletes this Telegram user's registered long-term memory across chats.

The host derives MemoryScope from Telegram's trusted user/chat ids. Model text cannot choose another user or tenant. The deletion registry stores only scope fields and opaque memory ids; conversation text stays in the vector store. The bot does not log tokens or message content by default.

Reproducible long-dialog check

The deterministic offline scenario inserts an access-code fact at turn 3 of a 100-turn conversation and gives FIFO and LRU baselines a capacity of 12:

python examples/telegram_long_dialog.py

Expected result: both bounded retention baselines lose the unaccessed early fact; semantic memory retrieves turn-2. This demonstrates a retention and retrieval difference, not literally infinite storage or guaranteed factual recall for arbitrary embedding models.