You now know enough to choose. Here is the decision.

When pure vector RAG is right, when you need a graph, and when hybrid earns its complexity.

This is the hinge lesson. Tier-1 of your mission was "understand the space well enough to decide." Everything from L1-L6 was loading the gun. This lesson is pulling the trigger: a framework you can apply to gbrain, Graphiti, or any client build.

The one idea: The choice is not "which is best," it is "what shape are your questions, and how fast does your data change." Match the machine to those two axes and the decision makes itself.

The three options, honestly

	Vector RAG	Knowledge graph	Hybrid (graph + vector)
Best at	"find passages like this"	"facts connected through relations"	both, at a cost
Setup cost	Low. Embed + store	High. Define entities, edges, extraction	Highest
Handles change	Re-embed, loses history	Temporal graph keeps history	Keeps history
Fails at	multi-hop, "as of when", exact ids alone	fuzzy "what does this remind me of"	over-engineering a simple need
Reach for it when	docs/FAQ search, one-shot Q&A	relationships + time matter	an agent that must reason over both

The two-axis test

Plot any KB need on two axes and the answer falls out:

Axis 1 · Question shapepassage-lookup → vector. relational/multi-hop → graph.

Axis 2 · Data velocitystatic corpus → batch is fine. live stream → temporal graph.

Low velocity + passage questions = plain vector RAG, do not overbuild. High velocity + relational questions = temporal graph (this is the agent-memory corner, where Graphiti lives). Mixed = hybrid, but only once a simpler build has visibly failed.

Applying it to your three goals

Need	Verdict
DuckyCutz client FAQ / KB (hours, services, barber traits)	Mostly static, passage-shaped. Vector RAG + BM25 is plenty. A graph here is over-engineering. (Lesson 1's ladder: do not build the graph you do not need.)
Your Rejig / personal brain (clients, deals, people, evolving)	Relational + high-velocity. Temporal graph. This is correctly Graphiti, or gbrain.
gbrain itself	Hybrid: vector + BM25 + RRF + graph signals, plus git for time. It spans both axes because a personal brain genuinely needs both.

The gbrain vs Graphiti verdict (the bet from L5-L6, resolved): Both target the high-velocity relational corner. gbrain bets on cheap deterministic linking + git-as-time to scale to 100k+ pages without LLM cost per write. Graphiti bets on LLM extraction + bitemporal edges for richer, time-precise relationships. Choose gbrain when volume and cost dominate; choose Graphiti when relationship richness and precise time-travel dominate. For Rejig you can run both: Graphiti as the live session memory you already have, gbrain as the bulk personal archive.

Retrieval practice

A client wants a static FAQ search bot. Right architecture?

Which two axes drive the architecture choice?

When does gbrain's bet beat Graphiti's?

Read this next (primary source)

★ Graph RAG: The Next Leap in Agentic Memory (AlphaMatch). The vector-vs-graph framing in one piece. ~10 min. Then re-skim the InfoQ hybrid article from Lesson 1 with new eyes.

Ask your teacher. This is the lesson to argue with. Try: "pressure-test my plan to use Graphiti + gbrain together", "talk me out of building a graph for DuckyCutz", or "what would I actually lose running only vector RAG for my personal brain?"

← Lesson 6: Temporal memory Next → Lesson 8: Synthesis and MCP (the answer, and the wire)