# Session 03 — Memory & Machinery · Facilitator Guide **Deck:** [deck-session-03.html](deck-session-03.html) · **Handout:** [session-03-exercises.md](session-03-exercises.md) · **Length:** 150 min, two acts **Session goal:** Act I — every student's agent has durable memory files that fix a real memory-gap miss from their week. Act II — every student can explain a real AI headline (memory prices, power deals, chip rules) using a correct mental model of the physical stack. This is the one session with a genuine lecture (the machinery act). It earns its place because it answers the question the whole course is secretly about: *why is any of this happening?* Keep it fast, concrete, and tied back to their own agent's cost. --- ## Run of show ### ACT I — MEMORY (0:00–0:55) #### 0:00–0:20 · Miss-log review (slides 1–2) - Students arrive with a week of misses. Slide 2: the three gap types (rule / memory / judgment). Teach the distinction with one example of each from the room, then have everyone **sort and count their own misses** into the three piles. - The reveal that drives the session: most rooms find ~70% are memory gaps — things no rule could fix because the agent never knew a fact about the student's life. *"Rules can't fix what the agent can't know. That's what tonight's first hour is for."* #### 0:20–0:35 · Why agents forget + the pattern (slides 3–5) - Slide 3, the desk-not-filing-cabinet metaphor. Keep it physical: "the context window is a desk that gets swept clean when you leave." - Slide 4, the memory/ folder pattern: index + one-fact files. Show your own memory folder on the projector if you have one — real examples beat the diagram. - **Slide 5 is non-negotiable content:** what never goes in memory. The three-part test (useful · harmless · yours). Say it twice. This is the privacy backbone and it gets tested by the partner audit next. #### 0:35–0:55 · EXERCISE 1 · Teach it who you are (slide 6) - Mine memory-gap misses → fact files. Ten is the target; five is fine. - Partner audit runs the three-part test on every proposed fact. Listen for facts that fail "harmless if leaked" — those are the teachable catches; surface one to the room (anonymized). - The recall test is the payoff: fresh session, re-run last week's hardest miss, watch it succeed. When it works, students visibly relax — "it remembered" is the emotional beat of Act I. ### 0:55–1:05 · Break (transition to machinery) - Use the break to write current pricing on the board (for Act II's worksheet) and the three fresh headlines you pulled in pre-flight. ### ACT II — MACHINERY (1:05–2:20) #### 1:05–1:10 · Act break (slide 7) - Reset the room's posture: *"we're going to leave your agent alone for forty minutes and go look at the building it lives in. When we come back, the news will read differently."* #### 1:10–1:35 · The lecture (slides 8–11) This is a real lecture. Pace it, but keep every claim concrete: - **Slide 8, the journey:** phone → tokens → fiber → the floor → back, under a second. Land "the typing effect is manufacturing in real time" — it recontextualizes something they've seen a hundred times. - **Slide 9, why GPUs:** chef vs ten thousand line cooks. Don't go deeper than "matrix math is thousands of identical multiplications; GPUs do thousands of identical things at once." - **Slide 10, memory bandwidth — the payoff slide.** 8 TB/s vs 0.05. Connect it to their wallet: *"the memory factories retooled for this, ordinary RAM got scarce, your next laptop costs more. That headline you saw? This is why."* This is the slide people quote back to you months later. - **Slide 11, the buildings:** 100,000+ GPUs = one job; gigawatts = cities; the why = agents, not chatbots. Callback to the pitch page's closing line. - Numbers are order-of-magnitude durable; if a student challenges a specific figure, agree cheerfully — "the exact number moved since I made this; the ratio didn't, and the ratio is the lesson." #### 1:35–1:55 · EXERCISE 2 · Price a thought (slides 12–13) - Calculators out, pairs. They find token counts in their OWN overnight log and compute what their night shift cost. - The reveal (slide 13): it's less than the coffee. Have a few read their number aloud. The gasp is real and earned — they computed it themselves. - Second half of the exercise matters as much: *"at that price, what else in your week is worth delegating?"* Three candidates each — these seed the post-course "next three agents" on the S04 closing slide. #### 1:55–2:20 · EXERCISE 3 · Read the headlines (slide 14) - Pairs draw a headline card, prep a 60-second no-jargon explanation, deliver to the room. Room votes "survives the skeptic?" - This is the session's assessment in disguise: if a pair can explain a real headline correctly using tonight's model, Act II worked. - ☆ SKIPPABLE if long: cut to 4 headline cards instead of 8, one per two pairs. Never cut it entirely — delivering the explanation is where the understanding sets. ### 2:20–2:30 · Homework + close (slides 15–17) - Add the "propose new memory facts" line to the nightly job (proposals only; student approves each morning). - Bring to S04: one surprising thing the agent remembered + miss-log delta (did misses drop after memory night?). The delta is real data the students generate about their own system — treat it as such. - Preview S04 honestly: *"next week is the most important session and the most fun. You attack each other's agents. Bring the one you trust."* --- ## Common failure points | Symptom | Cause | Fix | |---|---|---| | Memory file with a password/account number in it | Student conflated "useful" with "everything" | Stop, delete it together, re-teach the three-part test. Do it kindly and publicly-ish — everyone learns from the catch | | Recall test fails (fresh session doesn't use the fact) | Fact not in the index, or index not read at startup | Check MEMORY.md lists the file; check the startup instruction reads the index | | A student rejects the whole machinery act as "not for me" | Feels like a physics class | Redirect to their own cost number — "this is why your night shift is cheap; that's not physics, that's your budget" | | Pricing worksheet stalls | Can't find token counts in the log | `[INSTRUCTOR: know exactly where token usage appears in your runtime's logs and demo it]` | ## What good looks like at the door Durable memory files that pass the recall test · the three-part privacy test internalized (a fact was rejected during the audit) · a self-computed night-shift cost · a headline explained to the room's satisfaction.