AI Context Memory Persistence

Specific Problems

AI frequently encounters memory loss during interactions: it fails to recall prior information at the start of new sessions, forgets user preferences after topic switches, and repeats past errors. This "today told, tomorrow forgotten" behavior fragments conversations, requiring users to repeatedly provide context and reducing interaction continuity and efficiency.

How the Skill Works

Core Principle: AI memory must rely on file persistence, not just session context. This skill uses a structured file system to ensure important information is automatically loaded when new tasks begin.

• Task Classification and Triggers: Before processing any user message, the skill must output a @task_classify marker to strictly classify the task type. New tasks (e.g., large topic jumps, task type changes) trigger file reads; if it's a continuation task, reads are skipped to save resources.
• Memory File System: Key files include SOUL.md (storing user preferences), MEMORY.md (daily logs), .learnings/ERRORS.md (error records), and .learnings/LEARNINGS.md (correction logs). For new tasks, these files are read in sequence to restore full context.
• Writing Rules and Principles: When encountering user corrections, errors, or configuration changes, the skill must write immediately to the corresponding files. It adheres to principles of writing immediately (to avoid forgetting), writing specifically (e.g., recording "user is called xiaolong" instead of "has preferences"), categorizing storage (different information in different files), and including dates (for traceability).
• Execution Flow: Classify task → Read relevant memory files → Execute task → Immediately save important information to files. For instance, when a user corrects an error, it is logged to .learnings/LEARNINGS.md right away.

Use Cases and Considerations

This skill is suitable for:
- Conversational environments with frequent topic switches, ensuring related background is loaded each time.
- Multi-task parallel processing, relying on files to isolate contexts between tasks.
- Long-term project collaboration, where important rules are persisted to MEMORY.md for direct use in new sessions.
- Learning from errors to avoid repetition via ERRORS.md.

Important Considerations:
- Strict Task Classification Enforcement: No operations are performed without the @task_classify marker to prevent confusion between new and old tasks.
- Avoid Delayed Writes: Important information must be written to files immediately; do not rely on "writing later," as it may be forgotten once the context ends.
- File Maintenance Required: Log files like memory/YYYY-MM-DD.md can accumulate, so periodic archiving or cleanup is recommended to keep the system tidy.
- Reliance on Companion Skills: Paired with self-improving-agent, it provides more standardized formats and writing guidelines for .learnings/ files.

By implementing file-driven memory management, this skill enables reliable persistent storage, significantly enhancing conversation continuity and accuracy.