memory-governance

name	memory-governance
description	记忆治理 — 管理分层记忆系统的 Promote/衰减/去重/凝练，确保记忆质量与一致性
version	1.0.0
license	MIT
metadata	{"author":"blackplume233","actant-tags":"memory,governance,knowledge-management,curator","layer":"kernel","agent":"actant-curator"}

Memory Governance

Overview

Manage the layered memory system: Promote, decay, deduplication, and condensation. Ensure memory quality and consistency across Instance, Template, and Actant layers.

Instance Memory Promote

Purpose: Elevate high-value Instance memories to Template or Actant layers for broader reuse.

Promote Workflow

1. Candidate review: Identify memories with high confidence, repeated use, or cross-session relevance
2. Conflict resolution: Before Promote, check for contradictions with existing Template/Actant memories
3. Promotion: Move approved candidates to ac://memory/template/{templateId}/ or ac://memory/actant/learnings/
4. Source attribution: Preserve origin Instance ID for traceability

Conflict Handling

• If Instance memory contradicts Template memory: flag for human review or apply recency/confidence rule
• If multiple Instances have conflicting memories: arbitrate by confidence score, usage count, or manual merge

Memory Deduplication

Purpose: Reduce redundancy and maintain a single source of truth.

• contentHash detection: Compute hash of normalized content; detect duplicates across layers
• Merge strategy: Keep the most comprehensive version; merge metadata (confidence, timestamps, sources)
• Refinement: Produce a condensed version that preserves all unique information

Memory Decay (Expiration)

Purpose: Archive or retire low-confidence or stale memories.

• Confidence decay rules: Apply time-based or usage-based decay to confidence scores
• Archive threshold: When confidence < threshold, move to archive (read-only, excluded from retrieval)
• Eviction: When archive exceeds capacity, delete lowest-value entries
• Audit trail: Log all decay/archive/eviction actions for debugging

Memory Conflict Arbitration

Purpose: Resolve contradictions when multiple Instances have conflicting memories.

• Detection: Compare memories on same topic/entity; flag semantic contradictions
• Arbitration strategies:
  • Recency: Prefer newer memory
  • Confidence: Prefer higher confidence
  • Source authority: Prefer Template > Instance
  • Manual: Escalate to human when automated resolution unclear
• Merge: When possible, merge into a nuanced single memory ("Usually X, but in context Y, Z")

Cross-Template Generic Learnings

Purpose: Identify and elevate learnings that apply across multiple Templates.

• Recognition: Detect patterns that recur across different Template contexts
• Promotion target: ac://memory/actant/learnings/
• Use case: System-wide best practices, shared domain knowledge, common pitfalls
• Governance: Curator reviews before Promote; avoid Template-specific details in Actant learnings