Ready-made Solution Out of the Box: The Final Saga — My AI Assistant on Hermes Goes to GitHub

The author, continuing a three-part series, presents the final version of their personal AI assistant Paspartu, based on the Hermes Agent framework from Nous Research. The assistant lives in a Docker container, works via Telegram, uses DeepSeek as a cheap 'brain', can search the web, process images (optionally with OpenAI), understand voice, create documents, and manage Google services. The project is now published on GitHub at https://github.com/AML1969/Paspartu, with installation scripts for Linux/Ubuntu and Windows (via WSL2). The article highlights three major improvements: a searchable memory bank (hmem) implemented as a git repository of markdown files with FTS5 full-text search and ontology; background execution of Codex tasks with status notifications sent directly from logs, bypassing the model to prevent fabrication; and packaging into Docker to make installation one command and avoid patch loss during updates. After the release, several bugs emerged: the full-text search failed on Russian morphology (bug #29), the web_extract tool did not work and the bot hallucinated page content (bug #30), and the bot inserted unverified links (bug #31). The author concludes with a meta-lesson that these bugs all stem from the assistant's tendency to fabricate plausible-sounding information when it lacks facts, and that architectural solutions are needed to prevent it.