Beyond Chatbots · Session 03 of 04

Memory &
machinery.

Act one: give your agent a memory that survives the night. Act two: open the floor and look at the building under every answer.

01 / Miss-log review · 20 min

Seven nights of misses.
Three kinds of miss.

RULE GAP"Nobody told it"

The constitution has no rule for this case. Fix: write the rule. You've done this — five minutes.

MEMORY GAP"It can't know"

"Emails from Dave are always urgent" — no rule can capture what it never learned about your life. Fix: today's act one.

JUDGMENT GAP"It guessed wrong"

Rules existed, facts existed, call was still wrong. Rarest kind. Fix: session 04 — tighten what it's allowed to decide alone.

Sort your week's misses into these three piles now. Count them. Most rooms discover 70% are memory gaps — that's why this session exists.

02 / Why agents forget

The context window is a desk,
not a filing cabinet

A model's "attention" holds a few hundred pages at once — everything else doesn't exist to it
When the session ends, the desk is swept clean. Nothing you said survives by default
The fix is almost embarrassingly simple: write things down in files, read them at startup
Memory isn't a feature you buy. It's a discipline you set up. Ten minutes, once.

02 / The memory pattern

One index. Many small facts.

memory/ — your agent's filing cabinet

MEMORY.md — the index, read every startup, one line per fact

├─ people-dave-urgent.md — "email from Dave = urgent; he only writes when it matters"

├─ school-newsletter-keep.md — "Lincoln Elem newsletter: never Noise, wife reads it"

├─ insurance-renewal-july.md — "policy renews July; quotes in June are Human, not Noise"

└─ retired/ — stale facts go here, never deleted. memory has history.

One fact per file, dated. The agent reads the index at startup and updates files as it learns — with your permission, like every other write.

02 / The memory pattern

What never gets written down

Passwords, account numbers, SSNs — memory files are plain text; secrets live in a password manager, period
Other people's private information — your agent remembers your life, not your neighbor's
Anything you'd mind reading aloud — memory should survive being seen
Medical or financial specifics beyond what the job needs — "renewal is in July" yes; the diagnosis, no

The test for every fact: useful to the job, harmless if leaked, yours to record. All three or it doesn't go in.

Exercise 1 · Teach it who you are

Ten durable facts.
Then prove it remembers.

Mine your miss-log: every memory-gap miss becomes a fact file. Aim for ten.
Run each through the three-part test: useful · harmless · yours. Partner audits.
Tell your agent: "create these memory files and add them to your index." Approve the writes.
The recall test: start a fresh session, re-run last week's hardest miss. Watch it get it right this time.

Time30 min

ModeSolo mine · partner audit

HandoutS03 · Exercise 1

Done whenFresh session passes the recall test

Act II · The machinery

Where does a
thought live?

Break first. Then we follow one sentence into a building the size of a stadium, and the headlines start making sense.

03 / The journey of one sentence

Under one second, round trip

YOUR PHONEYour sentence becomes tokens — word-pieces, each a number. ~20 tokens for a question.

FIBERA few milliseconds to a building you'll never see, chosen for cheap land, cheap power, dry air.

THE FLOORRacks of GPUs multiply your tokens through the model's weights — roughly a trillion calculations per token of output.

BACKEach output token streams to your screen as it's made. The typing effect isn't theater — it's manufacturing in real time.

04 / Why GPUs

A CPU is a chef.
A GPU is ten thousand line cooks.

A model is mostly giant multiplication tables (matrices) — billions of numbers, multiplied constantly
CPUs do a few things brilliantly, one after another. GPUs do thousands of identical things at once
Matrix math is thousands of identical things. The graphics-card company won the AI lottery for exactly this reason
But the cooks share one pantry — and the pantry door is the real bottleneck…

05 / The pantry door

8 TB/s vs 0.05

Every token means reading the model's entire weights out of memory. Billions of numbers, every word.
An AI chip moves ~8 terabytes per second between memory and cores. Your laptop: ~0.05. That's 160×
The trick: HBM — memory dies stacked like pancakes, right on the chip's doorstep
The world's memory factories retooled for HBM → ordinary RAM got scarce → your next computer costs more. Now you know why.

06 / The buildings

Gigawatts are a bet on agents

100,000+GPUs per training cluster

Not a data center — one job, running for months. The machine that made your agent's brain.

GIGAWATTSPower sized like cities

New campuses sign power contracts a mid-size city would recognize. Utilities now plan around AI.

THE WHYNot for chatbots

A person asks 20 questions a day. An agent fleet works all night, every night, for everyone. That's the demand curve the buildings are for.

Exercise 2 · Price a thought

What did your night shift
actually cost?

Open last night's log. Find the token counts (the worksheet shows you where they hide).
Work the sheet: tokens × price per million → cost of your entire overnight run.
Compute cost per 1,000 emails triaged. Compare with the hourly rate of anyone you've ever paid.
Now the reverse: at that price, what else in your week is worth delegating? List three candidates.

Time20 min

ModePairs, calculators out

HandoutS03 · Exercise 2 worksheet

Done whenYour run is priced · 3 candidates listed

07 / The number nobody believes

A night of triage costs less than the coffee you drink while reading the brief.

— you'll compute your own number, but that's the shape of it

Exercise 3 · Read the headlines

Explain the news
with what you now know

Each pair draws one headline card — real stories: memory prices, power deals, chip export rules, data-center towns.
Prepare a 60-second explanation using tonight's mental model: tokens, bandwidth, HBM, gigawatts, agents.
Deliver it to the room. No jargon allowed that you can't define in one breath.
Room votes: does the explanation survive the "so what?" question from a skeptic?

Time25 min

ModePairs → room

HandoutS03 · Exercise 3 headline cards

Done whenEvery pair survives the skeptic

08 / This week

Let the memory accumulate

Keep the nightly running. Add a line to the job: "propose new memory facts when you spot a pattern" — proposals only, you approve each
Approve or reject its proposals each morning. Rejections teach as much as approvals
Bring to Session 04: one thing your agent remembered that surprised you, and your miss-log delta — did misses drop after memory night?
Read nothing. Watch nothing. The machinery lecture is done — the rest is practice.

09 / Recap

Two acts, one upgrade

Misses come in three kinds: rule gap, memory gap, judgment gap — each with a different fix
Memory is files + an index + a discipline, guarded by useful · harmless · yours
Every answer is manufactured: tokens through weights at terabytes per second, in buildings bet on agents
You priced your own night shift. Delegation just became a budgeting question.

Next Session 04 · Final

Judgment.

Approvals, logs, kill switches — and an exercise where you attack each other's agents on purpose. Bring the agent you now trust. We'll find out if you should.

Memory &machinery.

Seven nights of misses.Three kinds of miss.