HiveMem Archive Skill

Overview

This skill guides the agent in systematically persisting knowledge into the HiveMem long-term memory. It extracts facts, decisions, and context from the session.

When to Use

• End of session: When the user says "archive", "save", or "done".
• Significant discovery: When a root cause is found or a feature is implemented.
• Architectural decision: When a choice (e.g., "PostgreSQL over ChromaDB") is made.

Workflow

1. Analyze Session: Summarize the core achievement and its rationale.
2. Classify (Realm/Signal/Topic):
   • Realm: Top-level (e.g., projects, private).
   • Signal: Category (e.g., software, finance).
   • Topic: Specific topic (e.g., hivemem-auth).
3. Store Cell: Pass dedupe_threshold (e.g. 0.92) to hivemem_add_cell so the server encodes once and refuses to insert if a near-duplicate exists. Fill all L0-L3 layers.
   • L0: Content (Verbatim)
   • L1: Summary
   • L2: Key Points
   • L3: Insight (The "Why")
4. Extract KG Facts: For every atomic fact (Subject-Predicate-Object), call hivemem_kg_add.
   • Use on_conflict="return" on hivemem_kg_add to detect contradicting active facts without inserting.
   • If conflict: Invalidate old fact with hivemem_kg_invalidate before adding new.
5. Establish Links: Link the new cell to related knowledge via hivemem_add_tunnel.

Security & Transparency

• Bounded Scope: This skill only uses hivemem_* tools and the current project context.
• Human-in-the-Loop: Agents must use the agent token role, which sends all writes to the pending queue for admin approval.

---