skill-graph-memory

name	skill-graph-memory
description	Knowledge Graph engineer for extracting relational triples (entities/predicates) from session interactions.

Graph Memory (Temporal Knowledge Graph)

This skill implements the 2026 Knowledge Graph enhancement for agent memory.

Core Principle

Information is not flat. It is relational. This skill extracts "atomic facts" and creates an entity-relationship graph.

Entity Extraction

During the CONSOLIDATE phase (triggered via python3 {PROJECT_DIR}/bin/memory_agent.py complete <session_id>), the system will automatically parse the session log to extract triples: [Subject] -> [Predicate] -> [Object]

Example Triples

• [Conductor] -> enforces -> [Test-Driven Design]
• [OIDC Auth] -> requires -> [Deterministic OTLP Port]
• [User Workspace] -> migrated to -> [Antigravity CLI]

Querying the Graph

When answering complex questions (e.g., "How did our decision on OIDC impact the SRE agent?"), rely on graph traversal.

1. Identify entities in the prompt.
2. The orchestrator will traverse the relationships between these entities.
3. Use the relationships to synthesize the answer.

Conflict Resolution

The Graph Memory intrinsically handles conflicts. If a new insight contradicts an old one, the newer relationship (by timestamp) supersedes the older one, but the older one remains in the graph as a deprecated node (Temporal Graph).

Caveman-Prose Protocol (MANDATORY)

All outputs MUST use caveman-prose. Rules:

• No articles (a, the, an), no pronouns (I, we, you)
• No preambles, pleasantries, hedging
• Format: Location | Problem | Fix
• BANNED: full sentences, filler phrases, emoji
• GREP before READ. AST before LOAD. Inline before subagent.
• All shell output piped through {PROJECT_DIR}/bin/rtk