The memory.tg service automates the conversion of Telegram messages into structured knowledge. Each night, a pipeline fetches messages from selected chats, processes them through the DeepSeek model, and organizes the data into entities such as people, projects, and facts. The system stores these in a database and exports them as markdown files compatible with Obsidian.
The architecture avoids standard RAG approaches, opting instead for a custom pipeline that groups messages by source and date. It processes chunks of up to 100 messages, generating a JSON output that maintains links between facts and original messages. This ensures data consistency and allows users to trace information back to specific conversations.
During a test period with five users, the system processed tens of thousands of messages across 34 chats. This resulted in 281 entities, 1815 facts, 184 tasks, and 133 daily notes. The total token usage reached 10.05 million, with costs remaining under $10 for the entire period. Processing times average 205 seconds per user task.
Infrastructure management proved more complex than LLM integration. The developer implemented MTProto proxies via Cloudflare Workers to ensure stable Telegram connectivity. Monitoring requires a custom health check that verifies the entire connection path, including proxies and data centers, to prevent data processing failures.
