DEV Community
Follow
If You Disable Codex’s SQLite Logs, What Keeps Your Project Memory Alive?
Codex has been criticized for excessive diagnostic logging to a local SQLite database, causing performance issues like rapid SSD writes and slower responses. An unofficial workaround involves using a SQLite trigger to ignore new log entries, reducing disk activity but potentially obscuring diagnostic information. This workaround highlights a critical distinction between an agent's internal logs and the persistent "project memory" needed for continuity. Agent logs contain low-level operational details, while project memory requires a concise record of key decisions, verified evidence, and next steps.
This essential project continuity is not preserved by internal diagnostic databases, which are designed for the agent's self-diagnosis and are subject to changes in schema and format. The article introduces QiJu, a local-first record layer designed to capture this continuity context for AI coding agents like Codex. QiJu intentionally records only the essential facts—ground truth files, decisions, evidence, and next actions—allowing subsequent agents to resume work correctly without needing the entire historical session. This approach is auditable and handoffable, unlike information trapped within an agent's internal recall.
QiJu provides a practical solution for maintaining project continuity, especially when agent logs are disabled or unavailable. An upgrade command, `qiju update`, ensures that agent integration skills remain current across projects. The project registration process separates project location from host configuration, enabling efficient updates. While QiJu is in developer preview with limitations, its core purpose is to establish a deliberate, inspectable record of what matters for continuing AI-assisted development, a crucial distinction from internal diagnostic logging.
