Codex Turned Into a Personal Assistant With Memory on Markdown, Git, and Telegram

On Habr AI, a detailed case study was published about how Codex was transformed from a coding agent into a personal assistant with long-term memory. Instead of vector databases and complex RAG, the author built a manageable system based on Markdown, Git, roles, automations, and external sources like Telegram and Anki.

Why Not RAG

The author starts with a simple problem: a typical AI chat responds well in the moment, but struggles to live over time. Each new conversation loses part of the context, poorly distinguishes stable facts from temporary notes and hypotheses, and personalization remains locked within a specific thread. For a personal assistant, this is insufficient, because it needs not only to respond to questions but also to remember habits, goals, decision history, and recurring scenarios.

Instead of a typical stack with embeddings, vector DB, and GraphRAG, he chose a more transparent option: a hierarchy of Markdown files in Git. This approach, according to him, works better on a small to medium knowledge corpus that can be arranged into a clear structure. Files are human-readable, easily version-controlled, don't hide search logic behind a black box, and force you to think not only about retrieval but also about storage discipline.

The main idea of the article is this: in personal memory, architecture and rules matter more than the latest trendy technology.

How Jarvis Works

Jarvis is built on a repository divided into several layers: profile for stable user information, areas for main life spheres, events for 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 respond, and if necessary also update memory without cluttering the entire system.

Practical value emerges in specific scenarios that the author has already linked with Codex:

• English teacher role that relies on lesson history and can work with AnkiConnect
• Daily analysis of store discounts and assembly of a budget diet for mass gain
• Weekly reassembly of a sports plan based on current load and recovery
• Short summaries of tasks, unresolved decisions, and personal weekly results
• Connecting Telegram archives, where almost 300 thousand messages and over 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, analyst, operator of routine procedures, or wellness assistant. Thanks to this, responses become not generic, but tied to the user's real tasks.

Telegram here is needed not as an endless storage of raw data, but as a source for normalization: messages are indexed by year, voice messages are transcribed, person cards and general syntheses are updated based on dialogues. Without such a cleaning layer, memory quickly turns into a dump, and overly broad write permissions start to breed noise and duplicates.

"Each subagent is like an octopus in a separate aquarium."

What This Means

The article clearly shows where practical focus is shifting today around AI agents. For personal assistants and work Copilot scenarios, the winner will not be the one with a simply stronger model, but the one who better designs memory, roles, access boundaries, and maintenance costs. If this layer is made transparent and cheap to maintain, even an ordinary coding agent can start working as a truly useful digital assistant, especially where it is critical to remember the person and not lose context between sessions. This is exactly what makes this approach practical.