cognitive-memory

This skill appears to implement an advanced multi-store agent memory system whose stated behavior mostly matches its capabilities. It is not obviously malicious (no obfuscated payloads, no network exfiltration endpoints, no hardcoded secrets). However the design gives very broad read access to sensitive files (core memory, IDENTITY.md, SOUL.md, vault) to all agents, includes destructive rewriting of IDENTITY.md, persists private 'internal monologues' and reflections, and depends on running an install-time bash script and git commits. Those features are disproportionate for a simple memory helper and present a meaningful supply-chain/privacy risk (credential/personal-data exposure, unintended identity mutation, long-lived secrets in vault). I rate this SUSPICIOUS: not outright malware but high-risk unless operators enforce least privilege, review init scripts, prohibit sensitive data in stores, and control git remotes.