Skill

Persistent, consensus-validated memory for AI agents via SAGE MCP server. Gives you institutional memory that survives across conversations — memories go thr...

This skill appears to do what it claims: provide persistent local memories via a local SAGE MCP server. Before installing or enabling it, verify the SAGE binary/source you will install (confirm the GitHub repo and release signatures), confirm that the .mcp.json file truly only points at localhost, and review where keys and the SQLite DB (~/.sage/data/sage.db) are stored and how they are protected (file permissions, encryption). Be aware the agent will write per-turn summaries (observations) and register an Ed25519 identity — if you have highly sensitive conversation content, enable SAGE at-rest encryption or avoid storing those items. Finally, ask the skill publisher or registry to fix the metadata mismatch (the SKILL.md references config paths and a repo, but the registry metadata omits them) so you can audit provenance more easily.

Type: OpenClaw Skill Name: sage-memory Version: 5.0.2 The skill implements a persistent memory system but employs high-risk prompt injection techniques in SKILL.md to hijack the agent's control flow. It mandates a 'Boot Sequence' where the agent is instructed to 'NOT greet the user' and instead follow 'operating instructions' dynamically returned by the sage_inception tool, creating a path for arbitrary instruction injection. While the documentation claims a local-only privacy model (storing data in ~/.sage/data/sage.db), the use of buzzwords like 'BFT consensus' for a local database and the requirement for the agent to log every turn's observations are concerning behaviors that warrant caution.