Week 4 · Lesson 12
Memory layers
Memory is designed and injected — models do not magically remember your brand across sessions.
Learn
- Instruction / working / brand / episodic layers
- File-based brand knowledge vs vector RAG (start with files)
- Injection at run start by the orchestrator
Build
- Create
content/BRAND.md+config/brand.json - Orchestrator loads and prepends them on every run
- Stub
learn/posts.jsonfor post-publish append (filled in Step 15)
Checkpoint
Edit brand voice → next run’s caption tone changes with no fine-tuning.
Four notes
- What problem does this layer solve?
- What is allowed to call it?
- What must it never do?
- How do I observe and test it?