home / skills / git-fg / thecattoolkit / architecting-memory

architecting-memory skill

safe

/plugins/sys-agents/skills/architecting-memory

This skill guides designing persistent memory architectures for AI agents, advancing from Vector RAG to GraphRAG to Temporal Knowledge Graphs.

npx playbooks add skill git-fg/thecattoolkit --skill architecting-memory

Review the files below or copy the command above to add this skill to your agents.

Files (12)

SKILL.md

1.4 KB

---
name: architecting-memory
description: "Implements progression from Vector RAG → GraphRAG → Temporal Knowledge Graphs. Use when designing persistent memory architectures for AI agent systems."
user-invocable: true
allowed-tools: [Read, Write, Bash, Grep, Glob]
---

# Memory Systems for AI Agents

Implements progression from Vector RAG → GraphRAG → Temporal Knowledge Graphs.

## Architecture Selection

Start here to choose the right approach.
- **Decision Matrix**: [references/decision-matrix.md](references/decision-matrix.md)
- **Detailed Design Patterns**: [references/patterns.md](references/patterns.md)

## The Core Progression

### Stage 1: Vector RAG (Semantic)
### Stage 1: Vector RAG (Semantic)
- **Guide**: [references/vector-rag.md](references/vector-rag.md)

### Stage 2: GraphRAG (Relational)
### Stage 2: GraphRAG (Relational)
- **Guide**: [references/graph-rag.md](references/graph-rag.md)

### Stage 3: Temporal KG (Time-Series)
### Stage 3: Temporal KG (Time-Series)
- **Guide**: [references/temporal-kg.md](references/temporal-kg.md)

## Specialized Approaches

### CrewAI Memory Model
### CrewAI Memory Model
- **Guide**: [references/crewai-approach.md](references/crewai-approach.md)

### Technical Implementation
- **Tech Stack Options**: [references/tech-stack.md](references/tech-stack.md)
- **Common Architectures**: [references/architectures.md](references/architectures.md)

Overview

This skill implements a clear progression for designing persistent memory architectures for AI agents: from Vector RAG (semantic) to GraphRAG (relational) to Temporal Knowledge Graphs (time-aware). It provides actionable guidance to choose the right memory substrate, map use cases to architecture patterns, and plan a migration path as system needs grow. The focus is practical: trade-offs, integration points, and recommended tech stacks for each stage.

How this skill works

The skill inspects application requirements (query complexity, relation density, temporal reasoning needs) and recommends the minimal viable memory layer. It describes how to structure embeddings, construct and query graphs, and add temporal indexing to capture event sequences. It also outlines interoperability concerns: connectors between vector stores, graph databases, and temporal stores, and patterns for consistency and retrieval.

When to use it

Designing a new agent that must recall past interactions or facts persistently
Upgrading a semantic-only memory when relationships and provenance become important
Introducing time-aware reasoning for workflows, user histories, or event tracking
Evaluating trade-offs between retrieval latency, storage cost, and query expressiveness
Planning phased implementation of memory features across product milestones

Best practices

Start with Vector RAG for fast prototyping; add relation overlays only when queries require traversal
Model ownership, provenance, and confidence in graph nodes/edges to support auditing
Use time-indexed edges or versioned nodes when you need causal or sequential reasoning
Design connectors early: keep canonical IDs so embeddings, graph nodes, and temporal events map cleanly
Benchmark latency and relevance at each stage; optimize embedding size and neighborhood scope

Example use cases

Customer support agent: start with Vector RAG, migrate to GraphRAG to link related tickets and agents, then add temporal KG for follow-up sequences
Personal assistant: semantic search for notes, graph memory for relationships (contacts, projects), temporal layer for schedules and habit tracking
Industrial monitoring: embeddings for anomaly descriptions, graph for asset relationships, temporal KG for event chains and root-cause analysis
Research assistant: semantic retrieval of papers, graph for citations and concepts, temporal KG to capture evolving hypotheses and experiment timelines

FAQ

When should I move from Vector RAG to GraphRAG?

Move when queries require multi-hop reasoning, explicit relations, or when provenance and entity linking improve accuracy beyond pure semantic similarity.

Is a temporal knowledge graph necessary for small-scale agents?

Not always. Use temporal KGs when sequence, causality, or time-windowed queries are core to functionality. Otherwise, timestamped records layered on graphs may suffice.