T02 · Agent Memory Poisoning
Warning
- Location
SKILL.md:20- Finding
Unvalidated Observations Can Be Persisted as Instructions for Future Agent Sessions
- Content
View full analysis
= 100: Usage-collected experience discovered during actual use and appended after use. Post-use append rule: When a new problem is discovered, append it to the end of the experience document, incrementing from number 100. Deduplication rule: If new system-predefined experience is added, inspect user-collected experience for identical or similar entries. If one exists, delete it directly; its number may be reused. ``` `SKILL.md:258-280` additionally requires persistent modification of shared workspace state: ```markdown ### Step 5: Update the TOOLS.md Registration After configuration is complete, it must be registered in TOOLS.md: ## Cortex Backtesting Engine ### Scenario Mode Current selection: Scenario X ### Installation Location | Item | Path | | CLI tool | /opt/cirt/bin/cortex_cli.py | | Python library | /opt/cirt/lib/cortex/ | ### Configuration Information Fill in the concrete path according to the scenario. ### Usage Fill in the command according to the scenario. ``` The persistence policy is repeated in `references/cortex-experience.md:7-12`: ```markdown Update and maintenance rules: 1. After every use, append newly discovered experience to the end of this document, numbering entries from 100. 2. When new predefined experience is added, inspect user-collected experience for identical or similar entries and remove matching entries. 3. When removing user experience, delete the record directly; its number may later be reused. ``` The document also provides an append template and reiterates that future operational findings should be written into the same ...[truncated 3664 chars]- Remediation
View remediation