Context Memory: Persistent Memory for AI Agents

The Concrete Problem of AI Memory Loss

A common and frustrating issue in AI Agent development is the fracturing of context and memory. This manifests specifically as:
* Forgetting Between Sessions: Each new conversation starts with AI as if having "amnesia," unable to recall previous project backgrounds, user preferences, or prior commitments.
* Context Drop on Topic Switch: When switching discussions from Project A to Project B, the AI quickly forgets key details about Project A, leading to repetitive communication or errors.
* Inability to Iterate: After an AI makes a mistake, there is no recording mechanism, causing similar errors to recur across different sessions, preventing genuine "learning."
The core conflict lies in the fact that the AI's conversational context window is temporary and volatile, while the memory users expect is persistent and coherent.

How the Skill Works: Files as Memory

The "Memory Management" skill proposes an architectural solution: externalize the AI's memory into a structured file system and establish a strict protocol for reading and writing. It doesn't rely on the model's own "short-term memory" but instead has the AI actively manage its cognitive state by reading and writing to files.
Its core workflow follows the protocol specified in BOOTSTRAP.md, with key steps including:
1. Task Classification & New Task Judgment: Before acting on any user instruction, the AI must first perform @task_classify. It determines whether the current interaction is a "continuation task" or a "new task" by analyzing topic span, changes in task type, and the appearance of new keywords. This is crucial for avoiding redundant reads and improving efficiency.
2. On-Demand Reading of Memory Files: Once a new task is confirmed, the AI reads a specific set of documents in order:
* SOUL.md + MEMORY.md: Loads persistent personality and memory, such as user preferences and core rules.
* AGENTS.md + TOOLS.md: Loads workflows and tool configurations.
* .learnings/ERRORS.md + .learnings/LEARNINGS.md: Loads historical errors and correction experiences.
3. Execution & Immediate Saving: During task execution, any important information (like a user's correction or a newly discovered rule) must be immediately written to the corresponding file, adhering to the principle of "written down equals truly remembered." For instance, if a user corrects a method, the AI instantly updates .learnings/LEARNINGS.md.

Applicable Boundaries and Caveats

This skill is primarily suited for AI Agent scenarios requiring long-term, complex collaboration, such as personal assistants or project collaborators. Its effectiveness highly depends on:
* Strict Adherence to Protocol: The AI must unconditionally follow the BOOTSTRAP.md workflow, especially the task classification step; otherwise, the memory loading process will fail.
* Maintenance by Users/Developers: The file system requires a clear structure and ongoing content maintenance; otherwise, it may accumulate redundant or contradictory information.
* Not a Replacement for Model Capability: It solves the problem of cross-session state persistence, not the enhancement of single-session reasoning ability. The model's inherent "short-term memory" and "knowledge" remain the foundation.
In short, it provides a reliable set of read/write interfaces and operating procedures for an AI's "external memory."