#6 "The Usual, Mr. Thompson?"

AI memory is the difference between a first visit and a hundredth. The first time you use an agent, you explain everything — format, tone, structure, preferences. The hundredth time, the agent already knows. "Markdown format, summary under 3 lines, numbers in tables" gets applied without being asked. You stop repeating yourself. The agent starts anticipating what you need.

Restaurants figured this out long before AI did. The way a good restaurant treats a regular is the model for how AI memory should work.

First-Timer vs. Regular

A first-time guest walks into a restaurant. The server hands over the menu, explains the dishes, walks through the wine list. The guest deliberates, asks questions, orders. Good experience, but that's all it is.

A regular walks in, and everything changes. The manager recognizes them at the door. "Mr. Thompson! It's been a while!" On the way to the table: "Your usual window seat is ready." The server arrives. No menu. Instead:

"The tasting menu again today? That truffle risotto you mentioned last time — we just got fresh seasonal truffles. Want to try it? Chablis as always, or should we switch to a Barolo to match the truffle?"

Same restaurant. Same menu. Completely different experience.

What Makes a Regular a Regular

The restaurant remembering Mr. Thompson isn't magic. Someone wrote it down.

First visit: the server quietly checks. "You mentioned the truffle risotto — did you enjoy it?" A note goes in the book: "Prefers truffle risotto. Chablis, always." Next visit: "Likes light desserts" gets added. Third visit: "Window seat preferred." As this information builds, the restaurant learns Mr. Thompson. No need for the menu. Less explaining. Instead: "We got this ingredient in today — I think it's exactly your taste."

What makes a regular a regular is accumulated memory. Not one great night. Knowing a little more each time and adjusting a little more each time.

Three Layers

A restaurant's memory of its regulars has layers — and each one has an AI counterpart.

Preference memory. "Likes truffle risotto." "Drinks Chablis." "Prefers light desserts." Things you directly told them. Without this, they ask from scratch every time. In AI: "prefers reports in markdown," "wants bullet points, not paragraphs." You stop repeating yourself.

Pattern memory. "Orders tasting menus on weekends, single dishes on weekdays." "White wine in summer, red in winter." Things you never said — the restaurant noticed on its own. In AI: "wants summaries on Mondays, detailed reports on Fridays." The agent adapts without being told.

Context memory. "You mentioned it was your birthday last time." "You said you just got back from a business trip." Things that came up naturally in conversation. In AI: "mentioned a project deadline next week." Conversational context carries forward, and the agent can act on it.

As these three layers build, the guest increasingly feels: "this place knows me." The same thing happens with a well-configured AI agent.

When Memory Goes Wrong

Badly used memory is worse than no memory at all — in restaurants and in AI.

Mr. Thompson said "next time, go seafood-heavy." Six months later, they still bring out the truffle risotto. Trapped in old memory. The guest said "I've changed" and it wasn't reflected. In AI: the user says "stop doing it that way" and the old preference keeps applying. That's friction, not personalization.

Or memory that reveals too much. "Last time you came with that guest, right? The wine you had that night was..." That crosses from attentive to uncomfortable. Remembering what someone ordered is welcome. Remembering who they were with is intrusive. In AI: if the agent surfaces private details the user didn't ask about, trust erodes the same way.

Good memory — at a restaurant or in an agent — exists to be useful, not to impress.