Codex turned into a personal assistant with memory built on Markdown, Git and Telegram

Habr AI has published a detailed case study on turning Codex from a coding agent into a personal assistant with long-term memory. Instead of vector databases and complex RAG, the author built a controllable system on Markdown, Git, roles, automations and external sources such as Telegram and Anki.

Why not RAG

The author starts with a simple problem: an ordinary AI chat answers reasonably well in the moment but barely knows how to live across time. Each new conversation loses part of the context, struggles to tell durable facts from temporary notes and hypotheses, and personalization stays locked inside a single thread. That is not enough for a personal assistant, which must not only react to a question but also remember habits, goals, the history of decisions and recurring scenarios.

Instead of the typical stack of embeddings, a vector DB and GraphRAG, he chose a more transparent option: a hierarchy of Markdown files in Git. By his account, this approach works better on a small or medium knowledge corpus that can be laid out in a clear structure. The files are human-readable, easy to version, do not hide search logic inside a black box, and force you to think not only about retrieval but also about storage discipline. The article's central idea: in personal memory, architecture and rules matter more than the latest fashionable technology.

How Jarvis is built

Jarvis is built on a repository split into several layers: profile for stable information about the user, areas for the main spheres of life, events for the chronology, preferences for constraints and patterns, roles for response modes, skills for local procedures, assets for external sources, and inbox for raw material. The point is not the folders themselves but that the agent is given strict behavior: first search for relevant context, then answer, and when needed also update memory without cluttering the whole system.

The practical value shows up in concrete scenarios the author has already connected to Codex:

• an English teacher role that draws on lesson history and can work with AnkiConnect
• a daily review of store discounts and a budget meal plan for a mass gain
• a weekly rebuild of the workout plan based on current load and recovery
• short summaries of tasks, unfinished decisions and personal results of the week
• Telegram archives connection, where almost 300 thousand messages and more than 3 thousand voice messages have already accumulated

Special emphasis is placed on roles and automations. The same agent can act as a contextual life assistant, an analyst, an operator of routine procedures or a wellness assistant. As a result, answers are tied to the user's real tasks rather than being generic.

Telegram is used here not as an endless warehouse of raw data but as a source for normalization: messages are indexed by year, voice messages are transcribed, and people cards and general syntheses are updated from the dialogues. Without such a cleaning layer, memory quickly turns into a dump, and write permissions that are too broad start to breed noise and duplicates.

"Each subagent is like an octopus in its own aquarium."

What this means

The article shows well where the practical focus around AI agents is shifting. For personal assistants and workplace Copilot scenarios, the winner will not be whoever simply has the stronger model, but whoever better designs memory, roles, access boundaries and maintenance cost. If this layer is made transparent and cheap to maintain, even an ordinary coding agent can start working as a genuinely useful digital helper, especially where it is critical to remember the person and not lose context between sessions. That is what makes the approach practical.