Duplicate Flashcards: How to find them and which copy to keep
How to find duplicate flashcards by sorting, searching and exact-match finders, plus how to decide which copy inherits the review history.
If you only want the how: sort your deck by the question field so identical fronts land next to each other, search a rare word from the fact you suspect you wrote twice, and if your app has an exact-match duplicate finder, run it on the question field and then again on the answer field. That catches most of them in about ten minutes of scrolling.
The part nobody tells you is that finding them is the easy half. The hard half is that one of those two cards has a review history and the other one does not, and if you delete the wrong one you have just thrown away the only evidence your scheduler had about whether you actually know that fact.
My name is Daniel. I'm a third year engineering science student at UofT, specializing in machine intelligence, and I built Nebulearn, so I have an obvious conflict of interest whenever I talk about study apps. The mechanics below apply the same in Anki, Quizlet or anything else that schedules cards one at a time.
Here is the version of this that cost me a week. In second year I made cards from my thermo lecture PDF on a Sunday. On the Wednesday I read the textbook chapter in a browser tab and made more cards off the highlights. Both passes produced a card about the sign convention on entropy change. Slightly different wording, same fact. By the end of the month the Sunday card was sitting at a 21 day interval and the Wednesday card kept coming back every two days. I spent that week convinced I did not know entropy sign conventions. I knew the fact fine. What I did not know was the Wednesday phrasing.
A duplicate splits one memory across two schedules
Your scheduler does not know anything about facts. It knows about cards. Every card gets its own difficulty estimate, its own stability estimate and its own next due date, built only from the grades you gave that card. That is the whole point of a per-card algorithm, and it is why FSRS can hold one card at 60 days while its neighbour comes back tomorrow.
So when the same fact exists on two cards, three things happen at once. You review the fact roughly twice as often as the schedule intended, because two independent timers are both pointing at it. Each timer is working from half the data, because your successes on one copy were invisible to the other. And the second copy is scheduled as a brand new card even though the memory behind it is weeks old, which means it asks you the easiest question in your deck at the tightest interval in your deck.
That last one is what makes duplicates feel like a good thing at first. The new copy is easy, you press Good, it feels like progress. It is not progress. You are paying new card prices for a memory you already own, and every new card you add is a bill that comes due for months afterward.
How to find the exact duplicates
Exact duplicates are the ones where the text matches character for character, usually because you imported the same set twice or pasted the same line twice. These are worth clearing first because they are cheap to find and they are the ones that inflate your card count without adding a single fact.
Anki genuinely wins this round. It ships a duplicate finder in the card browser under the Notes menu, you point it at one field, and it tags everything that collides. It also flags a duplicate while you are still typing a new note of the same notetype, which stops the problem before it enters the queue. If you are in Anki, that is the first thing to run and there is nothing clever to add to it.
If your app does not have a finder, three manual passes get you most of the way:
- Sort the card list by the question field. Identical fronts become adjacent rows, and near-identical fronts usually land within a screen of each other.
- Sort by the answer field instead and scan again. This is the pass people skip, and it catches the case where you asked the same fact two different ways.
- Search for the distinctive noun. If you think you wrote the Nernst equation twice, search "Nernst" and read the six results.
Exact matching has two blind spots that matter more than the tool does. A trailing space, a different capitalization, or one copy that says "What is X" while the other says "Define X" will slip past any string comparison. And a cloze card like {{c1::glycolysis}} and a question-and-answer card testing the same word are stored as completely different text, so no finder will ever put them side by side. Clearing exact duplicates is maintenance. It is not the audit.
The near duplicates no finder will catch
Near duplicates are the expensive ones, and they are expensive for the reason my thermo cards were expensive: the two copies do not compete on wording, so both of them survive and both of them keep asking you for time.
A near duplicate is two cards that test the same retrieval. "Which enzyme catalyzes the committed step of glycolysis" and "Phosphofructokinase catalyzes which step" are the same memory in two directions. So are "What does the Chain Rule say" and a card with the formula on the front and the name on the back. So are a diagram card and a text card about the same label. No string comparison will find these, because there is no string in common.
The finder for near duplicates is the study session itself. When a card comes up and it feels like déjà vu, that is not a coincidence and it is not a sign you are learning. Tag the card right there. I use a tag called dupe-check and I look at it once a week, usually on a Sunday, which takes about five minutes because there are rarely more than a dozen.
The reason this works and a batch cleanup does not is that your queue already sorts your collection by how often you see things. A batch cleanup makes you read 1,400 cards to find nine problems. The queue hands you the nine problems for free over two weeks, because duplicated facts are exactly the facts you are seeing too often. If you are going to schedule a real sweep anyway, do it the way an exam deck audit works: bounded window, one deck, stop when the window ends.
Which copy keeps the review history
Once you have a pair in front of you, the instinct is to delete the worse one. That is half right. The card you want to keep and the wording you want to keep are often on different cards, and if you treat this as a delete decision you will lose one of them.
The move that solves it is to edit, not replace. Editing a card's text does not reset its schedule. So if the old copy has forty reps and clumsy wording, and the new copy has clean wording and three reps, you do not keep the new copy. You paste the new wording into the old card and delete the new one. You keep the schedule and the phrasing at the same time.
| What you are looking at | Keep | Why |
|---|---|---|
| Old copy, many reps, decent wording | The old one, delete the new | The interval is earned. Do not reset it for a cosmetic rewrite. |
| Old copy, many reps, bad wording | The old one, pasted over with the new wording | Editing preserves history. Deleting throws away forty grades. |
| Old copy at a long interval, new copy lapsing weekly | The new one, and start over honestly | The long interval was recognition, not recall. |
| Both copies young, neither much better | Either. Flip a coin and move on | Nothing is at stake yet. Deliberating costs more than the card. |
| Two copies you cannot merge, mid exam week | Neither. Skip or ignore one | Cleanup is not urgent. Getting it out of tomorrow's queue is. |
The third row is the one worth sitting with, because it is the only case where the old history is actively lying to you. If a fact is at 60 days on one card and failing every week on another, you have a card that taught you its own phrasing rather than the idea. That is an encoding problem wearing a duplicate costume, and the fix is the same as it is for any card you keep failing: rewrite it so it tests the idea, or apply the leech protocol and let it go.
When two decks should both keep the same fact
Not every pair is a mistake. Sometimes the same fact genuinely belongs to two courses, and the advice to merge is wrong.
The distinction I use is whether the two cards test the same retrieval or the same content. Bayes' theorem in a stats course and Bayes' theorem in a machine learning course is the same retrieval, so one card with two tags. The definition of entropy in thermo and the definition of entropy in information theory share a word and nothing else, so those are two cards and always were, and calling them duplicates would be a mistake.
For the genuine same-retrieval case, one card and two tags beats two cards in two decks, for a boring reason: your memory does not care which folder the card is in. Two cards means two schedules means you review a fact you know because it is cross-listed. One card with stats-mid2 and ml-final on it still lets you filter down to one exam's worth of material, and it does not double your daily load. This is also the quiet reason shared decks bloat so fast when you merge one into your own collection: you are not importing new facts, you are importing second copies of facts you already had.
When the cleanup is not worth it
Two situations where I stop. The first is a deck over a couple of thousand cards with a midterm inside two weeks. A dedup pass there is procrastination with a productive costume, and the honest fix is the one from maximum sustainable card count: decide what the deck is for, and use skip or ignore to get the second copy out of tomorrow's queue without spending the evening on housekeeping.
The second is moving between apps, and this one is worth being blunt about. Review history does not transfer across apps. Not from Quizlet, not cleanly out of Anki, not anywhere. So the "which copy keeps the history" question has exactly one answer during a migration, and the answer is neither of them. That means the order matters: dedup before you export, not after you import, because after the move every card is new and you have lost the only signal that told you which copy was the real one. If you are mid move, the specifics are in switching from Anki and moving Quizlet sets into a real scheduler.
The reason duplicates happen at all is that capture is scattered across a week. You make cards from a PDF on Sunday and from a browser tab on Wednesday and nothing connects the two passes. That is the part I built for: in Nebulearn, the PDF upload and the Chrome extension put both passes into the same collection, FSRS runs on the surviving card so its history is doing real work, and skip or ignore takes the loser out of the queue without deleting the fact or burning through the 1,000 saved questions on the free tier. It does not have a one-click duplicate finder, and if exact-match cleanup on a ten thousand card collection is your whole problem, Anki's browser is still the better tool for that hour. But if your problem is that the same fact keeps arriving twice because it arrived from two places, the fix is one place to put it and one schedule to trust.