AGI记忆模组

用户与模型之间的任何交互行为都可触发；提供Context Engineering五大核心能力（选择、压缩、检索、状态、记忆）及认知模型层支持；作为元技能强制常驻运行

This skill is a large, full-featured agent memory system and mostly does what its description says — but treat it as high-impact. Key things to consider before installing: (1) always: true means it will be force-resident and can run on every agent invocation — only install if you understand and accept that. (2) Review credential_manager, encryption, and credential_path logic: check how keys are generated/stored and whether any secrets could be written to disk or exfiltrated. (3) Verify the consent flow in PrivacyManager is actually enforced before any memory persistence occurs (SKILL.md delegates consent presentation to the agent). (4) If you plan to run it, sandbox it (isolated environment, limited file access), set storage_base_path to an audited directory, and ensure Redis/other backends are configured to trusted hosts only. (5) Prefer installing from a known, signed source or request provenance/maintainer details; if you cannot audit the code, do not enable it as always-resident. Providing the maintainer identity, a short audit of credential_manager/encryption code (how keys are stored and rotated), and confirmation that no external endpoints are contacted by default would increase confidence.

Type: OpenClaw Skill Name: memory-and-context-engineering Version: 1.0.10 The 'memory-and-context-engineering' skill bundle is a comprehensive and well-architected memory infrastructure for AI agents. It implements a sophisticated four-layer system (Orchestrator, Coordination, Storage, and Infrastructure) designed to manage agent context, cognitive modeling, and long-term memory. Key security features include a PrivacyManager (scripts/privacy.py) for user consent and sensitive data detection, a PermissionBoundaryController (scripts/permission_boundary_controller.py) for access control and data redaction, and a CredentialManager (scripts/credential_manager.py) that utilizes AES-256-GCM encryption for local storage. The code follows modern Python best practices, using Pydantic for data validation and standard libraries for infrastructure (Redis, Tiktoken). No evidence of malicious intent, data exfiltration, or unauthorized execution was found; the 'always: true' resident status is appropriate for its role as a foundational meta-skill.