Active recall: why it beats rereading
Elman Huseynov · 2026-08-25 ·
Rereading a solution is the weakest way to learn an algorithm. Rebuilding it from an empty page is the strongest.
The gap is large. It has been measured for over a hundred years.
Why rereading feels right
Read an explanation of Dijkstra twice. The second pass is smooth. Nothing surprises you.
Your brain reads that smoothness as knowledge.
What you built is recognition. The interview asks for recall.
Recognition skills come apart at the worst moment. An empty editor. Someone watching.
What the studies found
The effect has a name: the testing effect. There is a large literature behind it.
Roediger and Karpicke ran the standard version in 2006. Students read a passage. Then they either reread it or took a recall test.
Tested five minutes later, the rereaders won. Tested a week later, the recall group won by a wide margin.
Karpicke and Blunt repeated it in 2011 against concept mapping. That is a much more active kind of studying.
Retrieval practice still won.
The direction is consistent. Producing an answer beats reviewing one.
What recall looks like for an algorithm
Close every tab. Then, in order:
- Say what the algorithm does in three sentences, out loud.
- Write the core loop from memory. No notes, no autocomplete.
- Name one production system that runs it.
- Name the pattern it gets confused with, and the difference.
Open your notes and mark what you missed.
The marking is the whole exercise.
Failing, then seeing the answer, beats reading that answer twice.
It will feel worse
Recall is uncomfortable. You sit in front of a blank page and nothing arrives.
Robert Bjork calls this a desirable difficulty. The effort of retrieval is what makes the memory durable.
The session that felt bad is usually the one that worked. The smooth session taught you least.
That inversion is why most people never switch. The feedback points the wrong way.
One rule
Never read a solution twice.
Read it once. Close it. Rebuild it. Compare.
Recall also pairs with the spacing effect: the gap between attempts matters as much as the attempt.
Tracking the spacing effect takes counting two numbers.
How many problems you read last month. How many you could write today with the tab closed.
For most people the second number is a shock. It is also the only one an interview measures.
Where this is built in
AlgoPath makes recall the default. Every step ends with a gate.
Explain it with no notes. Rewrite the core from scratch. Say why it fits the real production case.
You cannot finish a step by reading it.
Proven steps are what the levels are built from.
Where recall sits in a week of preparation is in how to prepare properly.
You recall the pattern first, and that is trainable.
The first 33 steps are free.