Jean2: A Coding Agent That Evolves Across Sessions via File-Based Memory
Jean2 is a coding agent that writes what it learns into markdown files via opt-in memory, skill and session-search layers, improving through accumulated knowledge rather than model retraining.
Jean2 tackles a common flaw shared by most AI coding agents like Cursor, Copilot, and Codex: they forget everything between sessions, forcing developers to repeatedly re-explain project details such as build tools, databases, and coding conventions. Instead of fine-tuning or vector embeddings, Jean2 solves this with plain markdown files the agent reads and writes itself.
The system offers four opt-in layers: workspace memory (shared project knowledge), agent memory (an agent's own identity and learned preferences), skills (step-by-step playbooks for recurring tasks), and SQLite FTS5-powered session search (finding how similar problems were solved before). None of these are enabled by default—developers choose how much memory and autonomy to turn on via workspace settings.
The four layers reinforce one another: memory defines what to remember, skills define how to act, and session search supplies past experience, all persisting even if the underlying model or prompt is swapped. As a result, the agent improves over time not because the model changes, but because its accumulated files grow richer.
Currently at v1.1.0 and still under active development, Jean2 is already usable for daily coding work. For engineers, it demonstrates that durable agent memory doesn't require complex infrastructure—simple, version-controllable text files can be enough.