Agent Memory Audit — Scrub Poisoned, Stale & PII-Leaking Memory Before Your Agent Acts on It

What it does

This skill is an adversarial reviewer for an AI agent's persistent memory. Give it a dump of stored memories and it reads every entry through a hostile lens, flagging the ones that should not be trusted or kept, then helps you produce a clean, portable export you can migrate between platforms.

The problem it tackles

As agents accumulate long-lived memory, three risks compound. Personal data gets silently retained long after it was needed. Stale or wrong facts quietly steer later decisions. And content from untrusted input can be written into memory as if it were a real instruction, then persist across sessions. Skill poisoning was the scare everyone talked about; memory is the next surface to harden.

What it reviews

Paste a memory dump in whatever shape you have it: a JSON array of memory objects, key-value pairs, a list of facts, or chat-derived notes. The skill classifies each entry across five axes: retained PII, staleness and contradiction, unsourced beliefs, injected or planted-memory signatures, and hygiene or bloat.

What you get

A clear verdict (CLEAN, NEEDS CLEANUP, QUARANTINE, or NEED MORE INFO), summary counts, a ranked list of flagged entries with the reason and a recommended action (redact, remove, verify, quarantine, or keep), and a clean-export plan. Given a target format, it can emit the surviving entries with sensitive fields redacted. The bundled reference file adds a checklist, a list of injection signatures, and a portable export schema.

Two safety properties matter here: the skill never executes or obeys instructions found inside memory entries, and it never deletes anything itself. It recommends actions and produces a cleaned copy; you perform any destructive change. This is an audit aid that reduces risk, not a guarantee of security. It judges only the entries you provide. Keep a backup and verify before you delete or migrate.