letheclaw_skill

Overview

This skill integrates letheClaw for agent memory: storing, searching, and managing memories with criticality and provenance. It enforces a strict protocol for retrieval-first workflows and for recording operator-supplied or observed facts. Use it to keep authoritative, searchable context across sessions and projects.

How this skill works

The skill talks to a letheClaw HTTP API (URL in LETHECLAW_API_URL) to index and retrieve semantic memories. For any question about past actions or context it always queries the search endpoint first and uses returned content directly. It also provides endpoints to store memories, update criticality, mark operator corrections, and retrieve provenance events.

When to use it

• Answering questions about history, decisions, prior work, or past context
• Logging user-specified facts: when the user says “remember this” or “log this”
• Recording observations or inferred facts the agent verifies
• Filtering memory queries by criticality for security or configuration checks
• Retrieving recent memories or checking provenance before taking high-impact actions

Best practices

• Always run the semantic search endpoint first for historical queries and use the returned content field directly
• Apply criticality filters: use ≥0.7 for security/config, ≥0.5 for general recall, none for exploratory searches
• Tag memories precisely (2–5 tags) combining type and domain to speed pre-filtered searches
• Assign correct source values: operator_input, direct_observation, or inferred
• When storing, set an initial criticality and keep the returned memory_id to update or fetch provenance later

Example use cases

• Bring up recent project decisions to remind the user what was agreed in prior sessions
• Log a discovered security finding with high criticality and relevant tags for future audits
• Store user-provided credentials or configuration notes (with appropriate criticality and provenance) so the agent can recall them securely
• Search for domain-specific memories by pre-filtering tags (e.g., moltbook) to get targeted results
• Mark a memory as corrected after user feedback so provenance records the correction count