The Agent Skills Directory

[SAFE]: The skill is entirely self-contained, using only the Python standard library. It makes no network requests, does not call external APIs, and does not require third-party packages.
[CREDENTIALS_UNSAFE]: The skill includes a robust redaction engine in scripts/memory_core.py designed to protect against accidental credential exposure. It automatically identifies and masks various secret types, such as AWS keys, GitHub tokens, and private keys, using predefined regex patterns before data is persisted.
[DATA_EXFILTRATION]: To prevent local environment details from leaking into project files, the skill strips PII (such as OS usernames within file paths) from metadata when promoting memories to the committed tiers (L2/L3).
[INDIRECT_PROMPT_INJECTION]: The skill manages the risk of malicious instructions in session transcripts through a rigorous promotion pipeline:
Ingestion points: Data is extracted from local session transcripts (L0) in scripts/memory_extract.py.
Boundary markers: The skill relies on deterministic recurrence gates (multiple sessions and calendar days) and a 'Human-in-the-Loop' adoption process for long-term memory, effectively filtering out transient or adversarial injections.
Capability inventory: The skill's actions are restricted to reading and writing local configuration and state files within the project's .memory/ directory.
Sanitization: All content is sanitized through a redaction pass and blocked from promotion if sensitive data is detected, pending human review.
[SAFE]: All storage operations are performed with security in mind, including restricted file permissions (0600) and atomic file replacements to prevent data corruption or unauthorized access.

agent-memory