The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of durable rules and lessons from conversational data, which are then recalled in future sessions. This creates a surface for indirect prompt injection if malicious instructions are persisted and later treated as authoritative rules.
Ingestion points: Conversational history and explicitly proposed 'lessons' as described in SKILL.md are evaluated for promotion to memory.
Boundary markers: Memory cards use a specific structured YAML frontmatter format defined in references/memory-format.md to delimit metadata from the rule content.
Capability inventory: The skill allows the agent to write new memory cards to the file system (apply action) and read rules back into the active context (recall action).
Sanitization: The instructions in SKILL.md mandate checks for 'safe wording', 'evidence', and the exclusion of 'unsupported causal claims' and 'secrets' before a rule is promoted to memory.- [COMMAND_EXECUTION]: The skill requires the agent to execute a bundled Python script to resolve storage paths for memory cards.
Evidence: SKILL.md provides explicit command-line instructions: python <skill-dir>/scripts/storage.py --project-root <project> --kind memory --title <topic> [--scope global].- [DYNAMIC_EXECUTION]: The skill uses a local script to dynamically generate file paths and directory structures based on project metadata and user-provided titles.
Evidence: scripts/storage.py calculates project-specific keys using SHA-256 hashes of directory paths and performs slugification of titles to determine file system locations within the user's home directory.

agent-memory