memory-fabric_skill

Overview

This skill provides knowledge graph memory orchestration for entity extraction, query parsing, deduplication, and cross-reference boosting. It builds unified context from graph queries and returns ranked, de-duplicated memories suitable for RAG and agent decision-making. Use it to ensure no relevant memories are missed when designing memory orchestration.

How this skill works

The skill parses natural-language queries to extract intent and entity hints, then dispatches graph searches via mcp__memory__* endpoints. Results are normalized, deduplicated above a configurable similarity threshold, and augmented with cross-reference metadata when entities link across the graph. A final score combines recency, semantic relevance, and source authority to produce ranked results.

When to use it

• When you need unified context from multiple graph nodes for RAG or agent prompts
• When queries should consider entity relationships and multi-hop traversal
• When duplicate or overlapping memories must be merged and authority validated
• When you want cross-reference boosting for graph-backed evidence
• When building memory orchestration for chat agents or long-running sessions

Best practices

• Tune deduplication threshold (default 0.85) to balance recall vs. nuance
• Keep boost and max-results config in environment vars for runtime control
• Run parallel graph queries to reduce latency but handle partial failures
• Surface graph relations in metadata so downstream prompts can reason over provenance
• Fallback to recent memories when queries are empty or graph is unavailable

Example use cases

• Answering: 'What did database-engineer recommend about pagination?' with relationship context
• Extracting entities and relations from conversational logs to seed the knowledge graph
• Merging similar design notes from multiple authors and boosting cross-validated items
• Performing multi-hop traversal to gather broader context for a design decision
• Providing unified memory context at session start via a prompt hook