RAG & Memory

LLMs are stateless. Every conversation starts from zero, and every fact your model does not have is either a hallucination or a retrieval problem you have not solved yet. This is the day you solve it properly. You will chunk a real corpus four different ways and score each strategy against the same question set, build a FAISS index and choose its type on measured latency and recall rather than on a blog post, then give an agent a memory hierarchy and prove it works the only way that counts, clear the notes, run the same task again, and watch it repeat a step it used to skip. This was Day 2 of the two-day Context Engineering intensive and now stands on its own. You do not need Day 1. If you already have the foundations, start here.

Prerequisites

- **Python you can read and write**, you will be editing working scripts, not starting from blank files - **API access**, Claude Pro, Cursor Pro, or an OpenAI/Anthropic key - **Roughly how a model works**, tokens, embeddings, prompting - **Helpful but not required:** Context Engineering (Day 1). This class is standalone, if you already have the foundations, start here.

What you leave able to do

• design a packaged assistant over your documents
• design a recursive summariser for oversized documents
• re architect an agent as a stateless reducer
• reconcile an agent's run with a tool that failed
• select among agent architectures
• synthesise a memory hierarchy for an agent
• synthesise an agent that carries notes across its own runs

How you show it. Paired runs of the same multi-step task with the note store kept and cleared, where the cleared run repeats a step the noted run skips, plus the note text the agent wrote and the later turn that cites it.

It sits on one path

See what your hours buy.