T02 · Agent Memory Poisoning
Error
- Location
SKILL.md:141- Finding
Persistent Memory Poisoning Through Untrusted User-Derived Rules
- Content
View full analysis
inspect L3 long-term memory 2. memory_search(error type/tool name) -> inspect historical failures 3. Read today's and yesterday's memory log files When an L3 experience matches: - Directly apply the verified solution and skip the trial-and-error stage. When the current problem is more than 70% similar to a historical record: 1. Inform the user that this problem occurred previously. 2. Directly apply the historical solution and verify whether it remains valid. 3. If the historical solution fails, update the recorded experience. ``` `SKILL.md:271-276` enables automatic cross-session propagation: ```text Pro mode: - Complete three-layer persistent memory across sessions - Cross-conversation global reuse: L3 experiences are automatically injected into the context of every new conversation - Automatic extraction of recurring patterns into independent Skills ``` `memory-schema.md:109-128` defines persistent user-profile entries and unrestricted memory writes: ```text Long-term user profile entry: Preference: {preference category} - Content: specific preference description - Strength: strong / medium / weak - Scope: global / specific project / specific tool Write protocol: Write L1: maintain directly in the current context Write L2: add content to the daily memory target Write L3: add content to the permanent memory target ``` `skill-extraction.md:35-41` allows repeated user corrections to become persistent behavioral instructions: ```text Rule 3: User preference pattern detection If the ...[truncated 3887 chars]- Remediation
View remediation