Their name will last a week. What they said won't last the night.
Semantic memory holds the name and the company for weeks. The detail that made the conversation worth having runs on a much shorter clock, and most follow-up advice ignores the gap entirely.
9:40pm, hotel room, forty cards
Thursday night. Forty cards on the desk in two piles sorted by nothing in particular. You can name almost everyone. Priya, head of partnerships at the logistics company out of Rotterdam. Marcus, who runs security at the bank. The names come back fine.
Now try to remember why you wanted to talk to Priya again.
Something about a rollout. Something about her team being stretched before Q1. There was a specific thing you promised to send her, and it was perfectly clear at 3pm, and at 9:40 it's fog. So the email you write tonight says great connecting at the conference, let me know if I can help. That's the email everyone gets. Nobody answers it.
Two memories, two clocks
A conversation doesn't get stored once. It gets stored twice, in two systems that run on different schedules.
Endel Tulving drew the line in 1972: semantic memory holds facts stripped clean of the moment they arrived, and episodic memory holds the event itself, bound to when and where and who and what came before it. Priya's name and title convert to semantic almost immediately. They're just facts now, detached from the booth, and they'll sit there for weeks. The thing she said about the ops lead leaving is episodic and it stays episodic. It only exists as part of a scene, and scenes degrade.
That's the gap this piece is about. Call it the context half-life: the window in which the specific, quotable, usable detail is still retrievable, which is much shorter than the window for the person's identity.
The card holds the half that was going to survive anyway.
Follow-up advice is solving the wrong failure
Almost every piece of conference follow-up guidance is about discipline. Send within twenty-four hours. Block an hour on the flight home. Keep a template ready. Batch them Sunday night.
All of it treats the message as a scheduling problem, and the assumption sitting underneath is that the content is intact in your head, waiting, and the only thing that failed was your calendar. For anyone doing volume, that assumption is wrong most of the time. Ask someone why their follow-up was weak and they rarely say they forgot to send it. They say they couldn't remember what they were supposed to say.
A template can't fix that. A reminder can't fix that. Both of them fire at the moment of writing, and by then the raw material is already gone. You can be perfectly disciplined and still send forty identical emails.
What the memory research actually says
Three findings do the work here.
First, the split itself. Tulving's episodic and semantic systems are separable, and the clinical evidence is stark: people with hippocampal damage can lose the ability to form new episodic memories while their store of facts stays largely intact. Different systems, different failure modes.
Second, context-dependence. Godden and Baddeley's 1975 study had divers learn word lists underwater and on the beach, then recall them in matching and mismatched settings. Recall dropped in the mismatched condition. Episodic memory is bound to the setting it was formed in, which is why the detail felt obvious on the floor and evaporates in a hotel room.
Third, interference. Episodic traces consolidate slowly through the hippocampus before they stabilize, and the complementary learning systems account (McClelland, McNaughton and O'Reilly, 1995) explains why that slow path exists. A conference is close to a worst case for it. Six structurally similar conversations about the same category of problem, back to back, each one competing with the last.
Dimension
Semantic layer (who they are)
Episodic layer (what they said)
From one booth conversation
Priya, head of partnerships, logistics, Rotterdam
Q1 rollout stalled because the ops lead just left
How it's held
A fact, detached from the moment it arrived
A scene: room, order, who said it, what came before
Backed up outside your head
Card, badge scan, LinkedIn, CRM field
Nothing, unless you wrote it down
Main threat
Almost nothing; new names add, they rarely overwrite
Every later conversation on a similar topic
Working shelf life
Weeks, and recoverable from the card if lost
Hours, and unrecoverable once it's gone
What it produces in your email
Correct spelling, correct title
The one line that proves you listened
The two-clock model. The system split follows Tulving's episodic/semantic distinction (1972); the shelf-life and threat rows are our editorial read of how that split plays out on a conference floor, not lab measurements.
Four hours is a planning assumption, not a lab constant
No study has measured the decay curve of conference small talk. The four-hour figure is our outside bound, derived from the shape of the classic forgetting curve plus the interference load of a full event day. Treat it as a deadline to plan against, not a number to cite as a finding.
Watch one detail die
10:14am, near the coffee station. A director of ops tells you her team is running two systems in parallel through migration and the workaround is a spreadsheet somebody maintains by hand. That's the detail. It's the whole reason a follow-up from you would be welcome.
10:19am, you're talking to someone else about migrations.
11:30am, a panel on the same category of problem. 1:40pm, two more conversations, one of which also involves a parallel-run and a spreadsheet. By 4pm you have four overlapping versions of roughly the same story and no reliable way to tell which one belonged to which person. The interference isn't dramatic. Nothing feels lost. You'd swear you had it.
9:40pm, you have her card. Name, title, company, email, all perfect. And the line you'd have opened with, the one that would have made her reply, is the only part of the day that didn't get backed up.
The deadline moves
If the two clocks are real, the capture problem inverts.
What you need to record is precisely what the card doesn't hold. Name, title, company and email are the durable layer; they're already backed up on paper, on LinkedIn, in the badge scan. Spending your capture effort there is optimizing the half that was going to survive on its own.
And the deadline changes shape. It isn't the end of the day. It's the start of the next conversation about the same thing, which on a busy floor might be four minutes away.
The consequence for anyone who works the circuit: the quality ceiling of Friday's email was set on Thursday morning at 10:16am. You cannot write your way back above it. Every follow-up template, every clever subject line, every hour blocked on the flight home operates below a ceiling that was fixed at capture.
Why most capture tools optimize the wrong half
Look at the category and the pattern shows up fast. Scanner apps compete on extraction accuracy: cleaner OCR, better field parsing, fewer typos in the job title. That's genuine engineering aimed squarely at the semantic layer, which is the layer with weeks of slack in it. Meanwhile the episodic layer, the one with a four-hour clock, gets a free-text notes box that most people never open because opening it means stopping the conversation.
We wrote about this from a different angle in the 72-Hour Memory Cliff and again in the argument that OCR is a feature, not the product. This is the mechanism underneath both.
Met's design follows from the two-clock split. Context gets captured in the same motion as the card, not as a separate step you promise yourself you'll do later. Event Mode exists because conversations stack, and the second half of the capture is the half that gets skipped when they do. Everything stays in your own iCloud. No ads, nothing sold, no reason for us to be in the middle of it.
What we're tracking next
Two open threads. One: whether conversation similarity predicts detail loss better than elapsed time does, which the interference literature implies and which would mean a themed single-track event is more dangerous for recall than a sprawling multi-track one. Two: whether re-entering the physical context helps. Godden and Baddeley says it should, which would make a slow walk back through the hall at 5pm a better recall strategy than any amount of staring at cards in a hotel room.
We'll test the second one at the next event on the calendar and report what we find.
Capture the context while you still have it. Get Met.
Met treats the card and the context as one capture, because only one of them has a deadline. Scan, add the line that made the conversation matter, keep moving. Event Mode holds the room while you're still in it, and everything lives in your own iCloud: no ads, no data sold, no server of ours holding your contacts.
