← back to The Seams Are The Point

There Is No Delete

Episode 2 · The Seams Are The Point — transcript

Cold open

Nora: Here is the conclusion, first, because this episode earns the right to lead with it.

Desmond: In humans, every mechanism that looks like deletion turns out, on inspection, to be a change in accessibility.

Nora: Not one of them removes anything.

Desmond: That's not a slogan. That's the reading of about ninety years of memory research. And it means we've been building the wrong primitive.

Intro

Nora: This is The Seams Are The Point. I'm Nora.

Desmond: I'm Desmond. Episode two of five, and the source itself calls this the act that matters most.

Nora: Last time we found out capacity isn't the variable everyone assumes. This time, delete isn't either.

Iron rusts, but not because of time

Desmond: Start with the oldest wrong idea. Decay theory. The notion that memories just fade with the passage of time, like a photograph left in the sun.

Nora: Sounds intuitive.

Desmond: It sounded intuitive to everyone until nineteen thirty-two, when McGeoch demolished it with a line about iron.

Desmond: Unused iron may rust. But oxidation is responsible. Not time.

Nora: So the metal doesn't degrade because a clock is running. It degrades because something is happening to it during that interval.

Desmond: Exactly. Forgetting tracks what happens during the interval, not the length of the interval. Time is not the mechanism.

Nora: Then what is?

Desmond: Underwood, nineteen fifty-seven. He went back and reanalyzed the classic forgetting studies.

Desmond: Most of the lab forgetting wasn't decay at all. It was proactive interference. Old material blocking new.

Desmond: And here's the part that should embarrass the field a little. Twenty-four-hour recall, in the field's own baseline data, varied hugely.

Desmond: It depended on how many lists a subject had already learned before the one being tested.

Nora: So the canonical forgetting curve was partly just measuring how cluttered a given subject's head already was.

Desmond: Substantially measuring its own subjects' accumulated history. Not the passage of time at all.

Nora: Say that as an engineering sentence.

Desmond: Age-based TTL eviction is the discredited law of disuse, wearing a config file. A fact nobody contradicted and nobody re-queried is not thereby wrong.

Nora: So eviction shouldn't be driven by a clock.

Desmond: It should be driven by interference. A newer fact, a better-sourced fact, a fact that actually contradicts the one sitting in that slot. Not wall-clock age.

Desmond: Fair caveat, though. McGeoch refuted time alone, not decay in every form. There are real decay processes in contemporary neuroscience.

Nora: But the engineering point doesn't need decay to be false everywhere. It only needs the interference result to be true.

Desmond: Which it is.

Two strengths, one item

Nora: Okay, if it's not a clock, what's the actual model?

Desmond: Bjork and Bjork, nineteen ninety-two. The New Theory of Disuse. Every item you've ever learned has two separate numbers attached to it.

Nora: Two numbers.

Desmond: Storage strength — how well learned the thing is. And retrieval strength — how easy it is to get at right now. Those are not the same quantity.

Nora: Give me the illustration.

Desmond: A phone number you had for five years, two decades ago. High storage. Near-zero retrieval. You'd recognize it if someone read it to you, but you can't produce it cold.

Desmond: Your hotel room number on day three of a trip. The opposite. High retrieval, low storage — and you lose it the moment you check out.

Nora: So storage is basically permanent, and retrieval is the thing that actually moves around.

Desmond: That's the theory's own claim, and I want to be precise about how it's stated, because this is a modeling assumption, not a measured fact.

The Record — the axiom

The Record: Bjork and Bjork's own words: storage strength, once accumulated, is never lost — to at least a first approximation.

The Record: That is an explicitly stated modeling axiom. Not a measured constant. Report it that way.

Keep everything, govern access

Nora: This is where it gets useful. There's a running fight in the source between two database philosophies.

Nora: Datomic, which keeps everything forever. Datalevin, where deleted data is actually gone.

Desmond: And Bjork and Bjork dissolve the fight. Both sides are right — about different variables.

Nora: Storage is the Datomic argument.

Desmond: Keep everything. Storage is cheap, and it's append-only, so there's no engineering reason to ever throw a fact away.

Nora: And retrieval is the Datalevin argument.

Desmond: Govern retrievability aggressively, because it's zero-sum. Every superseded value that surfaces at read time costs a slot a correct fact needed.

Nora: So what's the actual threat to getting the current value back?

Desmond: Underwood again. The asymmetry. The biggest threat to retrieving the current value isn't age.

Desmond: It's the accumulated pile of previous values sitting under the same cue — including the ones you correctly kept, for audit.

Nora: So the audit trail is actively dangerous if it lives in the same place you query.

Desmond: Which gives us the rule. The audit log and the retrieval index have to be separate artifacts. History is retained. History does not get to compete at read time.

Nora: There's a second half of the Bjork assumption, right? The one nobody quotes.

Desmond: Right. Retrieval-strength loss from competitors is greater the higher an item's current retrieval strength, and the lower its storage strength.

Nora: Translate.

Desmond: Well-established facts resist being crowded out. So weight supersession by how established a fact is, not just by which one showed up last.

Desmond: A single low-confidence observation shouldn't bump a value that's been corroborated twenty times. Record it as a competing candidate, pending corroboration. Don't let it win by arriving second.

You cannot argue a memory away

Nora: So that covers old facts sitting quietly. What happens when you actively try to correct one?

Desmond: This is Mark Bouton's work. Extinction. Described in animal learning, decades before anyone had a language model to break.

Nora: Ghost memory.

Desmond: That's the mechanism behind it, yes. When you extinguish a conditioned response, you have not destroyed the original learning.

Desmond: The cue now has two available meanings, and context decides which one you get.

Nora: And the correction doesn't get equal footing with the original.

Desmond: It gets worse footing. The second thing learned is far more context-dependent than the first.

Desmond: The original keeps broad, default scope. The correction gets filed narrowly, tied to the context it arrived in.

Desmond: Two consequences follow, and both are directly applicable. First: you cannot retract by piling up contradicting evidence.

Desmond: That is literally the extinction protocol.

Nora: Meaning what, concretely?

Desmond: Bouton and Swartzentruber ran eighty-four extinction trials after eight conditioning trials. The response was gone after about twenty.

Desmond: Then they returned the animal to the original context, and it came straight back.

Nora: So more repetitions of 'no, that's wrong' don't make the correction stick.

Desmond: Rauhut, Thomas and Ayres, two thousand one, confirmed it directly. One hundred extinction trials produced no less renewal than twenty.

Desmond: Piling on more of the same does not work.

Desmond: Fair caveat, though — the same paper found two treatments that did block renewal. Explicitly-unpaired presentation, and conditioned inhibition.

Desmond: So 'nothing works' overstates it. It's specifically piling on more of the same that fails.

Nora: So in schema terms — retraction has to be a distinct, marked operation. Not just another row saying the opposite thing.

Desmond: Right. And that leads to the second consequence, which the source calls the highest-leverage, lowest-cost decision in the whole document.

Nora: That's a big claim. What is it?

Desmond: A correction stored as just another row that has to win a competition will lose that competition, in any context that isn't the one it was learned in.

Desmond: The fix comes from Brooks and Bouton's work on cues. Present a reminder of the correction at retrieval time. Store the retraction co-indexed with the original value's own cues.

Nora: So you can't fetch the stale value without also fetching the correction.

Desmond: Any lookup that surfaces the old value necessarily surfaces 'this was corrected, see X,' in the same fetch. That's the whole trick. One join, not a competition.

The Record — what could not be verified

The Record: A note on honesty. The two Bouton sentences most worth quoting in this section could not be verified verbatim against the paywalled original.

The Record: The substance is supported by the paper's abstract. They are paraphrased here, not quoted — and the design rule survives on the Debunking Handbook guidance, which was verified directly.

Retract by replace

Nora: So a correction needs to travel with the original. Does the correction's content matter too, or just its placement?

Desmond: Both. This is the continued influence effect. People keep using retracted information in their reasoning even after they've accepted the correction. It's well replicated.

Nora: Even when they explicitly know it's false?

Desmond: Even then. And the Debunking Handbook, two thousand twenty, is blunt about the fix. Don't rely on a simple retraction. Supply an alternative account with the same explanatory relevance as the original.

Nora: So the schema rule is—

Desmond: Retract-by-replace, never retract-by-null. A tombstone with no replacement leaves a gap, and downstream reasoning fills that gap with the stale value anyway.

Nora: What about the backfire effect? Corrections that ironically make people believe the false thing harder. That was the scary one.

Desmond: It largely failed to replicate.

The Record — the failure that's real

The Record: The backfire effect largely failed to replicate. Do not build machinery to prevent it.

The Record: Spend that effort on the failure that is actually real — corrections that are incomplete, narrowly scoped, or lost at retrieval time.

Soft delete, and forgetting as a feature

Nora: One more distinction before we wrap. Directed forgetting — telling someone to forget a list.

Desmond: It impairs free recall while leaving recognition of the same items largely intact. You tell them to forget it, and later they can't generate it — but shown it, they still recognize it.

Nora: Excluded from generative retrieval, versus absent from the store.

Desmond: Exactly that mapping. Biology suppresses the first and preserves the second. That's a well-evidenced default for soft delete.

Nora: Which means hard delete is the exception, not the default.

Desmond: It should exist only as a separate, explicitly destructive, separately-audited compliance operation. Not the everyday path.

Nora: Last piece. Why have any of this at all — why not just keep everything fully accessible forever?

Desmond: Richards and Frankland, twenty seventeen. Transience is a feature, not a bug. It serves flexibility — reducing the pull of outdated information on decisions.

Desmond: And generalization — preventing overfitting to specific past episodes.

Nora: So perfect retention isn't the goal.

Desmond: It's a design flaw if you actually hit it. And the rate shouldn't be one global number.

Desmond: Tune it per content class, to that class's measured volatility. An order's status changes hourly. A refund threshold changes yearly.

Outro

Nora: So let's collect the rules. Eviction runs on interference, not age.

Desmond: Storage stays forever and is cheap. Retrieval is governed aggressively and is zero-sum.

Nora: The audit log and the retrieval index are separate artifacts — history doesn't compete at read time.

Desmond: Retraction is a distinct, marked operation, co-indexed with the original's own cues, never a pile of contradicting rows.

Nora: And every retraction carries a replacement. Never a null.

Desmond: Which brings us back to the top. Nothing gets deleted. Accessibility changes. That's the whole architecture.

Nora: Next time — why one memory system was never enough. Two tiers, catastrophic interference, and the rat experiment that tells you what gets promoted.

Desmond: Episode three. Two Systems and a Gate.