The Agent Skills Directory

[SAFE]: The skill performs purely local operations, reading from MEMORY.md and writing state to the ~/.openclaw directory. No network-enabled libraries (like requests or urllib) or system-level command execution utilities (like subprocess or os.system) are used.
[SAFE]: Deserialization of state files is handled securely using yaml.safe_load(), which prevents the execution of arbitrary Python objects during file loading.
[SAFE]: The summarization logic is deterministic and implemented via string manipulation and regular expressions in Python. It does not utilize an LLM for the compaction process itself, which eliminates risks related to prompt injection during the generation of the DAG nodes.
[SAFE]: The skill's persistence via a cron schedule is a standard feature of the platform for maintenance tasks and does not involve unauthorized modification of system startup scripts or user profiles.

memory-dag-compactor