Lessons

12 lessons, roughly 5 hours if you do the labs.

Part 0 — Orientation

0. What this is, and what it isn't10 min
The honest version of the graph engineering pitch, minus the hype that got it trending.
1. Setup: Hermes and Copilot CLIHands-on15 min
Get one of the two drivers running, plus the lab repo. Ten minutes, no API key.

2. Graphs versus vectors, and why the answer is both20 min
Similarity is not structure. What each store is actually good at, and where the hybrid seam goes.
3. The temporal model: time is not metadataHands-on25 min
Put time on relationships, preserve expired facts, and answer questions about the world as it was.

4. Extraction, and what it actually costsHands-on25 min
Why ingestion dominates graph-memory cost, and how cached prefixes, low effort, and batching change the bill.
5. Agentic workflow graphs: prompts with contracts25 min
Build a directed system of narrow, testable nodes instead of one giant prompt.
6. The validation gate: stopping bad data before it compoundsHands-on30 min
A knowledge graph has a nasty property. Bad data doesn't just sit there, it reproduces.

7. The routing policy: make the tradeoff executableHands-on20 min
A model-routing policy that separates cheap ingestion from expensive, evidence-bound traversal.
8. Wiring the graph into Hermes and Copilot CLIHands-on30 min
Expose your graph over MCP so your agent has memory that survives the session. Includes a config from the viral article that does not work.
9. Measure it, do not assume itHands-on25 min
The metrics that expose extraction quality, cache behavior, retrieval bounds, and real cost per episode.

10. The production track: Neo4j and GraphitiHands-on30 min
When SQLite stops being enough, and the verified Graphiti MCP config the viral article got wrong.
11. What you actually learned15 min
The transferable lesson, the mistakes worth remembering, and how to read the next viral architecture thread.