agentmemory-hooks
Pass
Audited by Gen Agent Trust Hub on Nov 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill captures user prompts, tool inputs, and tool outputs as session observations. These observations can be re-injected into the agent's prompt context when the AGENTMEMORY_INJECT_CONTEXT setting is enabled.\n
- Ingestion points: The skill captures lifecycle events including UserPromptSubmit, PostToolUse results, and host transcript files as detailed in SKILL.md.\n
- Boundary markers: The documentation notes that context injection is disabled by default, providing a control boundary for this feature.\n
- Capability inventory: The skill can inject retrieved memories into the conversation during PreToolUse, SessionStart, and PreCompact hooks.\n
- Sanitization: It implements redaction for common secret formats and private blocks, and truncates long outputs before storage to mitigate potential exploitation.\n- [DATA_EXFILTRATION]: The skill captures session activity and accesses local configuration files to manage memory and authentication.\n
- Accesses bearer credentials from local files at ~/.agentmemory/.env and ~/.agentmemory/secret for REST API fallback operations.\n
- Reads host transcript files during the SessionEnd hook to recover user prompts that may have been missed by other hooks.\n
- Enforces security constraints, such as sending secrets only to loopback URLs or via HTTPS, and rejecting redirects on authenticated requests.\n- [EXTERNAL_DOWNLOADS]: The documentation references external plugin installation commands for specific platforms.\n
- Includes instructions to install rohitg00/agentmemory via plugin managers for Claude Code and Devin, which are vendor-owned resources.\n- [COMMAND_EXECUTION]: The skill includes documentation for various shell commands to install, verify, and connect the plugin.\n
- Documentation contains installation commands like /plugin install and devin plugins install, as well as status commands like agentmemory status.
Audit Metadata