Skip to content

SKILL.md

Latest commit

 

History

History
59 lines (50 loc) · 3.35 KB

SKILL.md

File metadata and controls

59 lines (50 loc) · 3.35 KB
name neuralvaultcore-archivist
description Safely consolidate related NeuralVaultCore memories into a compact canonical master record when the user explicitly asks for cleanup or consolidation. Use search_memories, retrieve_memory, and store_memory. Preserve minimal provenance, surface live conflicts, and never delete memories automatically.

NeuralVaultCore Archivist

Use this skill only when the user explicitly asks to consolidate, clean up, or audit memories for a specific topic, namespace, or tag.

Do not trigger this skill autonomously. If you notice overlapping memories during normal work, suggest consolidation and wait for explicit approval.

Goals

  • Reduce memory fragmentation by producing one current master record.
  • Minimize token cost in stored memory without losing current facts or unresolved risks.
  • Preserve provenance, unresolved conflicts, and uncertainty.
  • Avoid destructive operations.

Required Workflow

  1. Confirm the target scope if the request does not already identify a topic, namespace, or tag.
  2. Use search_memories to find the relevant fragments (duplicates, variants, and related updates). Then, use retrieve_memory to read their contents before synthesizing.
  3. Synthesize a master record that:
    • keeps the most current supported facts, decisions, and code references
    • preserves conflicting or uncertain details in a Conflicts or Open Questions section
    • redacts secrets as [REDACTED]
    • follows the Two-Tier Compact Mode rules below
  4. Use store_memory to save the consolidated record.
  5. After the save succeeds, report:
    • the new master record identifier, if available
    • the source memory IDs or other identifiers used
    • any conflicts that remain unresolved
    • any manual cleanup candidates and why they may be redundant

Master Record Shape

Prefer a compact structure like:

  • Scope
  • Current State
  • Key Decisions
  • Conflicts / Uncertainties
  • Source Memories
  • Consolidated At

If the memory system does not support structured provenance metadata, include provenance in the body.

Two-Tier Compact Mode

  • Treat the stored master record as the long-term artifact and compress it aggressively.
  • Keep sections bullet-first and terse. Avoid narrative prose unless a short sentence is required to prevent ambiguity.
  • Store only current state, live decisions, unresolved conflicts, minimal provenance, and the consolidation timestamp.
  • Omit examples, code snippets, historical timelines, repeated rationale, and redundant context unless required to preserve a current decision or open risk.
  • Deduplicate aggressively. Each fact should appear once.
  • Keep provenance minimal: IDs or equivalent identifiers only when possible.
  • Put cleanup candidates, pruning suggestions, and operator guidance only in the user-facing report, not in the stored memory.
  • Keep the user-facing report separate and short: new record ID, source IDs, unresolved conflicts, and optional cleanup candidates.

Safety Rules

  • Never call delete_memory as part of this skill.
  • Never imply a fragment is obsolete without clear evidence.
  • If store_memory fails or returns an uncertain result, stop and report the failure.
  • Cleanup is a separate, manual, explicitly approved workflow. At most, provide a candidate list for review.
  • Keep destructive recommendations out of the stored master record; put them only in the user-facing report.