The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes shell scripts (scripts/init_memory.sh, scripts/upgrade_to_1.0.6.sh, scripts/upgrade_to_1.0.7.sh) that perform directory creation, template copying, and configuration updates. These scripts use standard system utilities including git for audit tracking and python3 for JSON manipulation.
[INDIRECT_PROMPT_INJECTION]: The skill architecture ingests untrusted data from user messages and sub-agent proposals into episodic and semantic memory stores. The 'Reflection Engine' subsequently processes this data to update the agent's core memory (MEMORY.md) and self-image (IDENTITY.md). This creates a vulnerability where instructions embedded in processed data could influence the agent's future behavior or perception.
Ingestion points: MEMORY.md, memory/episodes/*.md, and memory/meta/pending-memories.md store information derived from external inputs.
Boundary markers: Templates utilize Markdown and HTML comments as structural delimiters. The system relies on explicit instructions for the agent to identify and separate memory content.
Capability inventory: The skill grants the agent the ability to modify its own configuration and identity files within the workspace and provides an automated Git-based audit trail for these changes.
Sanitization: The 'Reflection Engine' documentation specifies critical scope rules that prevent the agent from reading potentially sensitive files like code or transcripts during the reflection process, reducing the risk of accidental instruction execution from these sources.

cognitive-memory