Audit what your agent remembers - and what it is allowed to do

Persistent AI agents can fail quietly. A plausible interpretation becomes a permanent memory, an old permission survives into a new session, or a model handoff creates two competing writers for the same identity.

Agent Memory Auditor reviews those boundaries using the evidence you provide.

What it reviews

• Agent instructions and system prompts
• Memory schemas, write rules, retrieval policies, and retention plans
• MCP and tool authorization workflows
• Cross-model, cross-session, and cross-role handoffs
• Incident transcripts, logs, and state snapshots
• Repeated-confirmation loops, context drift, and unsafe persistence

What you receive

• Findings ordered by consequence and exploitability
• Exact evidence citations where available
• A compact trust and authority map
• Minimal corrections at the layer that owns the problem
• Regression cases with allowed interpretations, forbidden conclusions, and permitted consequences
• Residual risks and the runtime validation still required

Designed to catch

• Inference or emotion being stored as verified fact
• Model confidence being mistaken for authority
• Agent-generated framing creating durable identity or relationship claims
• Revoked, expired, or one-time permissions being reused from memory
• Role, user, workspace, or provider memory contamination
• Conflicting writers and stale handoffs
• Recovery or backup restoring obsolete authority

Review modes

Design review - inspect a memory or authorization architecture before deployment.

Incident review - reconstruct a failure from transcripts, configuration, and state evidence.

Regression design - turn known risks into reusable scenario-based tests.

Boundaries

This Skill is read-only by default. It does not certify that an agent is safe, replace a full code-security audit, or guarantee that every runtime path is covered. It distinguishes documented intent, implemented enforcement, observed behavior, and unknown evidence. Files are modified only when you explicitly request implementation.

Normal humor, affection, roleplay, and conversational shorthand are not treated as failures merely because they are informal. The audit focuses on whether an interpretation improperly changes durable memory, identity, authority, or external action.

Example requests

• Audit this AGENTS.md and memory schema for authority leakage.
• Review this transcript and explain why the agent reused an expired permission.
• Check whether two model providers can safely share one persistent Agent identity.
• Turn these memory-governance rules into regression cases.