T02 · Agent Memory Poisoning
- Location
setup.md:75- Finding
Persistent Agent-Memory Poisoning Through User-Controlled Rules
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This is a disclosed local memory skill, but it needs Review because it persistently stores free-form corrections and uses them to steer future agent behavior.
Install only if you want a persistent, cross-session agent memory that can influence future work. Review and approve exactly what it writes to memory and workspace steering files, avoid storing sensitive or third-party data, and do not enable the optional companion install unless you can verify its source and version.
setup.md:75Persistent Agent-Memory Poisoning Through User-Controlled Rules
setup.md:90Unpinned Third-Party Skill Is Installed Before Security Review
boundaries.md:44Complete-Deletion Workflow Creates a Residual Export Copy
learning.md:29Person-Specific Information Is Persisted Despite the Declared Third-Party Data Prohibition
The activation criteria in the metadata description are very broad, covering routine situations like failed commands, user corrections, noticing a better approach, or explicit reference to the skill. That can cause the skill to trigger in many normal interactions and begin persistent memory-related behavior without sufficiently clear user intent, increasing the chance of over-collection or unexpected state changes.
The 'When to Use' section uses vague natural-language triggers like noticing output could be better or wanting to evaluate significant work. In a self-modifying or memory-writing skill, ambiguous triggers are risky because they normalize autonomous invocation and can expand the skill's behavior into situations where the user did not clearly consent to persistence or follow-through.
The skill authorizes automatic promotion, demotion, and archival of behavioral patterns based on time and frequency thresholds, which is autonomous state management affecting persistent memory. Even though deletion requires asking, these unsupervised changes can still alter future agent behavior, preserve inaccurate inferences, or hide important context from the hot set without user review.
- Pattern used 3x in 7 days → promote to HOT
- Pattern unused 30 days → demote to WARM
- Pattern unused 90 days → archive to COLD
- Never delete without asking
### 4. Namespace Isolation
- Project patterns stay in `projects/{name}.md`
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
| Financial | Card numbers, bank accounts, crypto seeds | Fraud risk |
| Medical | Diagnoses, medications, conditions | Privacy, HIPAA |
| Biometric | Voice patterns, behavioral fingerprints | Identity theft |
| Third parties | Info about other people | No consent obtained |
| Location patterns | Home/work addresses, routines | Physical safety |
| Access patterns | What systems user has access to | Privilege escalation |
The phrase "forget everything" is a broad natural-language trigger for a destructive action, and the document does not constrain it to an explicit command channel, confirmation step, or scope. In a conversational agent, ordinary discussion, quoting, testing, or prompt injection could accidentally or maliciously trigger memory deletion, causing loss of retained state and disrupting agent behavior.
The kill-switch workflow performs export and wipe immediately after a trigger phrase without requiring an explicit warning that deletion is irreversible. This increases the chance of accidental destructive action and leaves users without a clear opportunity to cancel before data is removed.
The template explicitly instructs creation of persistent files under ~/self-improving/, including memory, corrections, and heartbeat state, without any user warning, consent step, retention limit, or privacy guidance. In a self-improving agent context, this is more dangerous because it normalizes long-lived local storage of potentially sensitive user-derived data and behavioral history, which can lead to unintended data retention or privacy leakage.
The skill specifies automatic loading of memory files at session start, including contextual and potentially personal preference/history data, without any user-facing notice or consent step. In a memory-heavy self-improving agent, silent retrieval of stored data can surprise users, expose sensitive prior context in unintended conversations, and undermine privacy expectations.
The correction workflow automatically persists user inputs across corrections logs, namespace files, and indexes, but does not clearly warn users that their messages may be stored in multiple locations. This increases privacy and data minimization risk because casual corrections, preferences, or project details may be retained and propagated more broadly than the user intended.
The setup explicitly directs creation and later modification of persistent files in the user's home directory and workspace, but it does not require informed consent or clearly warn that these changes are durable. In an agent skill context, this is risky because it establishes long-lived memory and steering files that can silently influence future behavior across tasks and sessions.
The maintenance routine automatically moves, archives, and compacts memory data based on age, yet gives no warning about retention periods or lifecycle transitions. While less immediately dangerous than collection itself, undisclosed archival can preserve data longer than users expect and complicate deletion, review, and compliance with user intent.