AGI记忆模组

This skill addresses the issues of AI agents easily forgetting context in long conversations, being unable to associate information across sessions, and lacking structured memory management. In practical development, developers often need to manually maintain conversation history, handle token budgets, and design retrieval logic. This skill encapsulates five core capabilities (Selection, Compression, Retrieval, State, and Memory) into a unified framework, automatically performing noise filtering, chain structure extraction, result re-ranking, task progress tracking, and conflict detection, enabling agents to truly possess "memory."

Usage is straightforward: simply set this skill as a meta-skill to run persistently (always: true), and all interactive behaviors will automatically trigger memory processing. You can then manage memory using natural language instructions, such as "Save the key points of the current conversation" or "Retrieve the architectural plan discussed last time." The underlying system employs a four-layer architecture, with the ContextOrchestrator serving as the central control hub for unified scheduling of token budgets, retrieval decisions, and multi-source coordination. The cognitive model layer provides dual-track parallel processing (semantic bucket refinement + chain extraction), ultimately storing results in the short-term/long-term memory layer.

It is particularly suitable for teams or individuals building complex AI agents, especially those already encountering context window limitations, needing cross-session knowledge reuse, or aiming to implement persistent memory for personalized assistants. AI agent developers, full-stack engineers, and AI product managers can all benefit, as it significantly reduces the engineering complexity of memory management, allowing developers to focus on business logic rather than underlying memory orchestration.

It is recommended to configure the storage path and Redis connection (optional) before first use and adjust token budget parameters based on actual scenarios. Note that this skill depends on Python 3.9+ and several third-party libraries (pydantic, tiktoken, etc.); ensure environment compatibility during deployment. As it runs persistently as a meta-skill, it will continuously consume a certain amount of tokens for memory compression and retrieval. It is advisable to regularly clean up or archive old memories in long-running tasks to optimize performance.