Memory Science
26 evidence-backed learning techniques drawn from 192 peer-reviewed papers, and the 13 memory principles behind them — mechanism, evidence, boundary conditions, and what each implies for a mnemonic-generation system.
This page is the evidence base behind the roadmap’s study aids — the spaced-repetition decks, audio recall cues and inline memory prompts used across the site all draw on the techniques catalogued here. Every technique links out to peer-reviewed papers, and each paper is DOI-verified so you can trace any claim back to its source. Techniques are ordered by how much published evidence supports them; the principles section below explains the mechanisms, boundary conditions and generation-system design implications behind them.
Choose a technique by goal
Start from what you’re trying to learn — each goal links to the techniques best supported for it.
Technique catalogue
26 techniques ordered by how much published evidence supports them; every paper is DOI-verified.
Cloze deletion (the Anki {{c1::…}} pattern) forces active recall of one or two specific facts inside an otherwise complete, true sentence. The evidence corpus supports this: spaced repetition through Anki tracked with higher Step-1 pass rates and lower failure rates (Cooper et al. 2023); self-generated Anki flashcard volume independently predicted licensing performance (Deng, Gluckstein & Larsen 2015); higher reliance on Anki active-retrieval cards tracked with better exam outcomes (Gilbert et al.
- Nawaf Salah Ayad Mohamed, Maria Abdulaziz Alrafi, Barbara Albani, Hassan Mohammed Abdu, Reem Albuhairan, Naziha Samir Altala, Sami Fatehi Abdalla — 20252025DOI: 10.1007/s40670-025-02487-5
- Juliana Magro, So-Young Oh, Nikola Košćica, Michael A. Poles — 20242024DOI: 10.1111/tct.13798
- Ogunjobi F., Alexander S. M., Cramer L. — **Year:** 20242024DOI: 10.7759/cureus.70994
- Michael M. Gilbert, Timothy C. Frommeyer, Garrett V. Brittain, Nickolas A. Stewart, Todd M. Turner, Adrienne Stolfi, Dean X. Parmelee — 20232023DOI: 10.1007/s40670-023-01826-8
- Jillian K. Wothe, Lindsey J. Wanberg, Rae D. Hohle, Aliya A. Sakher, Laura E. Bosacker, Faizel Khan, Andrew Olson, David J. Satin — 20232023DOI: 10.1177/23821205231173289
- Spencer Cooper, Nicole Twardowski, Michael P. Vogel, Daniel Perling, Rebecca Ryznar — 20232023DOI: 10.5195/ijms.2023.1549
- Cole P. Thompson, Marion A. Hughes — 20232023DOI: 10.1016/j.jacr.2023.08.028
- Dylan Jape, Jessie Zhou, Shane Bullock — 20222022DOI: 10.1186/s12909-022-03324-8
- David M. Harris, Michael F. Chiang — 20222022DOI: 10.7759/cureus.23530
- Cyrus Anthony Pumilia, Spencer Lessans, David Harris — 20202020DOI: 10.7759/cureus.10372
- Tabibian B., Upadhyay U., De A., Zarezade A., Schölkopf B., Gomez-Rodriguez M. — **Year:** 20192019DOI: 10.1073/pnas.1815156116
- Settles B., Meeder B. — **Year:** 20162016DOI: 10.18653/v1/P16-1174
- Tiago Taveira-Gomes, Rui Prado-Costa, Mílton Severo, Maria Amélia Ferreira — 20152015DOI: 10.1186/s12909-014-0275-0
- Francis Deng, Jeffrey A. Gluckstein, Douglas P. Larsen — 2015 (cited in the build brief as 2016; OpenAlex records 2015)2015DOI: 10.1007/s40037-015-0220-x
- Karpicke J. D., Roediger H. L. — **Year:** 20082008DOI: 10.1126/science.1152408
Retrieval practice / testing effect
15 papersRetrieval practice — actively producing an answer from memory before re-reading it — is the single best-supported memory technique in the corpus. Dunlosky et al. (2013) rate practice testing one of only two "high utility" strategies; Adesope et al.
- Shana K. Carpenter, Steven C. Pan, Andrew C. Butler — **Year:** 20222022DOI: 10.1038/s44159-022-00089-1
- Pooja K. Agarwal, Ludmila Nunes, Janell R. Blunt — **Year:** 20212021DOI: 10.1007/s10648-021-09595-9
- Chunliang Yang, Liang Luo, Miguel A. Vadillo, Rongjun Yu, David R. Shanks — **Year:** 20212021DOI: 10.1037/bul0000309
- Yana Weinstein, Christopher R. Madan, Megan Sumeracki — **Year:** 20182018DOI: 10.1186/s41235-017-0087-y
- Steven C. Pan, Timothy C. Rickard — **Year:** 20182018DOI: 10.1037/bul0000151
- Does retrieval practice enhance learning and transfer relative to restudy for term-definition facts?Steven C. Pan, Timothy C. Rickard — **Year:** 20172017DOI: 10.1037/xap0000124
- Olusola Adesope, Dominic A. Trevisan, NarayanKripa Sundararajan — **Year:** 20172017DOI: 10.3102/0034654316689306
- Bernhard Pastötter, Karl-Heinz T. Bäuml — **Year:** 20142014DOI: 10.3389/fpsyg.2014.00286
- Christopher A. Rowland — **Year:** 20142014DOI: 10.1037/a0037559
- Adam L. Putnam, Henry L. Roediger III — **Year:** 20132013DOI: 10.3758/s13421-012-0245-x
- John Dunlosky, Katherine A. Rawson, Elizabeth J. Marsh, Mitchell J. Nathan, Daniel T. Willingham — **Year:** 20132013DOI: 10.1177/1529100612453266
- Jeffrey D. Karpicke, Janell R. Blunt — **Year:** 20112011DOI: 10.1126/science.1199327
- Henry L. Roediger III, Andrew C. Butler — **Year:** 20112011DOI: 10.1016/j.tics.2010.09.003
- Jeffrey D. Karpicke, Henry L. Roediger III — **Year:** 20082008DOI: 10.1126/science.1152408
- Roediger H. L., Karpicke J. D. — **Year:** 20062006DOI: 10.1111/j.1467-9280.2006.01693.x
Spacing / distributed practice
13 papersA spacingPrompt is a short inline instruction telling the learner WHEN to come back to a chapter — not what to re-read, but when. The whole design rests on a few robust corpus findings:
- E. B. Murray, Aidan J. Horner, Silke Melanie Goebel — **Year:** 20252025DOI: 10.1007/s10648-025-10035-1
- Futing Zou, Brice A. Kuhl, Sarah DuBrow, J. Benjamin Hutchinson — **Year:** 20252025DOI: 10.1016/j.celrep.2025.115232
- David W. Price, Ting Wang, Thomas R. O'Neill, Zachary J. Morgan, Prasad Chodavarapu, Andrew Bazemore, Lars E. Peterson, Warren P. Newton — **Year:** 20242024DOI: 10.1097/acm.0000000000005856
- Marjolein Versteeg, Renée A. Hendriks, Aliki Thomas, Belinda W. C. Ommering, Paul Steendijk — **Year:** 20192019DOI: 10.1111/medu.14025
- Behzad Tabibian, Utkarsh Upadhyay, Abir De, Ali Zarezade, Bernhard Schölkopf, Manuel Gomez-Rodriguez — **Year:** 20192019DOI: 10.1073/pnas.1815156116
- C. Zerr, Jeffrey J. Berg, Steven M. Nelson, Andrew K. Fishell, Neil K. Savalia, Kathleen B. McDermott — **Year:** 20182018DOI: 10.1177/0956797618772540
- John L. Dobson, José Martí Pérez, Tracy Linderholm — **Year:** 20162016DOI: 10.1002/ase.1668
- Paul Smolen, Yili Zhang, John H. Byrne — **Year:** 20162016DOI: 10.1038/nrn.2015.18
- Shana K. Carpenter, Nicholas J. Cepeda, Doug Rohrer, Sean H. K. Kang, Harold Pashler — **Year:** 20122012DOI: 10.1007/s10648-012-9205-z
- Kornell N. — **Year:** 20092009DOI: 10.1002/acp.1537
- Cepeda N. J., Vul E., Rohrer D., Wixted J. T., Pashler H. — **Year:** 20082008DOI: 10.1111/j.1467-9280.2008.02209.x
- Nicholas J. Cepeda, Harold Pashler, Edward Vul, John T. Wixted, Doug Rohrer — **Year:** 20062006DOI: 10.1037/0033-2909.132.3.354
- Donovan J. J., Radosevich D. J. — **Year:** 19991999DOI: 10.1037/0021-9010.84.5.795
The keyword mnemonic (Atkinson & Raugh 1975) encodes an unfamiliar word in two links: an acoustic/visual keyword that sounds or looks like the target, and a vivid interactive image binding that keyword to the target's meaning. For SQL learners the "foreign vocabulary" is the reserved-word lexicon — HAVING, COALESCE, PARTITION BY, INTERVAL and friends — which are opaque exactly like L2 words: the surface form gives little hint of the operation. We therefore pick a familiar keyword that is acoustically/visually near the SQL token, then stage a one-scene image whose action is the term's actual semantics (so the image is gate-checkable against the docNote, not just a pun).
- Kejia Qu; Tianzhi Liu; Yihuan Qiao; Pengcheng Wang — 20242024DOI: 10.1016/j.heliyon.2024.e25212
- Jaewook Lee; Andrew Lan — 20232023DOI: 10.48550/arxiv.2305.10436
- Chia-Hui Chiu; C. F. Hawkins — 20232023DOI: 10.23971/jefl.v13i2.6313
- Toshiya Miyatsu; Mark A. McDaniel — 20192019DOI: 10.3758/s13421-019-00936-2
- Fritz C., Morris P., Acton M., Voelkel A., Etkind R. — **Year:** 20072007DOI: 10.1002/acp.1287
- Wyra M., Lawson M., Hungi N. — **Year:** 20072007DOI: 10.1016/j.learninstruc.2007.02.008
- Alvin Y. Wang; Margaret H. Thomas — 19951995DOI: 10.1037/0022-0663.87.3.468
- Alvin Y. Wang; Margaret H. Thomas; Judith A. Ouellette — 19921992DOI: 10.1037/0022-0663.84.4.520
- Mark A. McDaniel; Michael Pressley — 19841984DOI: 10.1037/0022-0663.76.4.598
- Raugh M., Atkinson R. — **Year:** 19751975DOI: 10.1037/h0078665
- Richard C. Atkinson; Michael R. Raugh — 19751975DOI: 10.1037/0278-7393.1.2.126
- Atkinson R. — **Year:** 19751975DOI: 10.1037/h0077029
The corpus is overwhelmingly about lab TMR: pairing material with an odor or sound at encoding, then re-playing that cue during monitored, often closed-loop sleep. A learner reading /postgres cannot run this. So the aids deliberately draw only on the naturalistic, actionable kernel these papers share:
- Judith Nicolas, Bradley R. King, David Lévesque, Latifa Lazzouni, Gaëlle Leroux, David Wang, Nir Grossman, Stephan P. Swinnen, Julien Doyon, Julie Carrier, Geneviève Albouy — 2025 (Nature Communications)2025DOI: 10.1038/s41467-025-57602-2
- Julia Carbone, Susanne Diekelmann — 2024 (npj Science of Learning 9)2024DOI: 10.1038/s41539-024-00244-8
- Leila Salvesen, Elena Capriglia, Martin Dresler, Giulio Bernardi — 2024 (Sleep Medicine Reviews 74:101908)2024DOI: 10.1016/j.smrv.2024.101908
- Dan Denis, Jessica D. Payne — 2024 (eNeuro 11:5)2024DOI: 10.1523/eneuro.0285-23.2024
- Dominique Recher, Judith Rohde, Giulia Da Poian, Mirka Henninger, Luzius Brogli, Reto Huber, Walter Karlen, Caroline Lustenberger, Birgit Kleim — 2024 (Translational Psychiatry 14:490)2024DOI: 10.1038/s41398-024-03192-4
- Mahmoud E. A. Abdellahi, Anne C. M. Koopman, Matthias S. Treder, Penelope A. Lewis — 2023 (eLife 12)2023DOI: 10.7554/eLife.84324
- Christine Barner, Ann-Sophie Werner, Sandra Schörk, Jan Born, Susanne Diekelmann — 2023 (Frontiers in Sleep)2023DOI: 10.3389/frsle.2023.1187170
- Hu X., Cheng L. Y., Chiu M. H., Paller K. A. — **Year:** 20202020DOI: 10.1037/bul0000223
- Rasch B., Born J. — **Year:** 20132013DOI: 10.1152/physrev.00032.2012
- Diekelmann S., Born J. — **Year:** 20102010DOI: 10.1038/nrn2762
- Rudoy J. D., Voss J. L., Westerberg C. E., Paller K. A. — **Year:** 20092009DOI: 10.1126/science.1179013
Text-to-speech / audio learning
11 papers- Simantiraki O., Wagner A. E., Cooke M. — **Year:** 20232023DOI: 10.3389/fnins.2023.1235911
- Singh A., Alexander P. A. — **Year:** 20222022DOI: 10.1007/s10648-021-09653-2
- Clinton-Lisell V. — **Year:** 20222022DOI: 10.3102/00346543211060871
- Clinton V. — **Year:** 20192019DOI: 10.1111/1467-9817.12269
- Singer L. M., Alexander P. A. — **Year:** 20172017DOI: 10.3102/0034654317722961
- Rogowsky B. A., Calhoun B. M., Tallal P. — **Year:** 20162016DOI: 10.1177/2158244016669550
- Reinwein J. — **Year:** 20112011DOI: 10.1007/s10936-011-9180-4
- Moreno R., Mayer R. E. — **Year:** 20072007DOI: 10.1007/s10648-007-9047-2
- Ginns P. — **Year:** 20052005DOI: 10.1016/j.learninstruc.2005.07.001
- Moreno R., Mayer R. E. — **Year:** 20022002DOI: 10.1037/0022-0663.94.1.156
- Ralston J. V., Pisoni D. B., Lively S. E., Greene B. G., Mullennix J. W. — **Year:** 19911991DOI: 10.1177/001872089103300408
AI-assisted / adaptive learning
10 papersThese are illustrative auto-generated quiz questions of the kind an automatic question generation (AQG) system would produce, each tied to a Postgres chapter and carrying evidence provenance. They model the JSON shape the lane emits (see spec §JSON shape). Each is a recall-or-recognition item with a correct answer and, where useful, plausible distractors — the multiple-choice form recommended by the AQG review (Kurdi et al.
- Hongming Li, Salah Esmaeiligoujar, Nazanin Adham, Hai Li, Rui Huang2026
- Maity S., Deroy A., Sarkar S. — **Year:** 20252025DOI: 10.1016/j.caeai.2025.100370
- Younes-Aziz Bachiri, Hicham Mouncif, Belaid Bouikhalene2025DOI: 10.34105/j.kmel.2025.17.018
- Suhana Bedi, Yutong Liu, Lucy Orr-Ewing, Dev Dash, Oluwasanmi Koyejo,2025DOI: 10.1001/jama.2024.21700
- Benjamin Paddags, Daniel Hershcovich, Valkyrie Savage2024DOI: 10.18653/v1/2024.bea-1.29
- Wanyong Feng, Jaewook Lee, Hunter McNichols, Alexander Scarlatos, Digory Smith,2024DOI: 10.18653/v1/2024.findings-naacl.193
- Taimoor Arif, Sumit Asthana, Kevyn Collins-Thompson2024DOI: 10.1145/3657604.3664714
- Semere Kiros Bitew, Johannes Deleu, Chris Develder, Thomas Demeester2023DOI: 10.48550/arXiv.2307.16338
- Ghader Kurdi, Jared Leo, Bijan Parsia, Uli Sattler, Salam Al-E'mari2020DOI: 10.1007/s40593-019-00186-y
- Wang Z., Lan A. S., Nie W., Waters A. E., Grimaldi P. J., Baraniuk R. G. — **Year:** 20182018DOI: 10.1145/3231644.3231654
Desirable difficulties
10 papersRereading and "I've seen this, it makes sense" feel like learning but are weakly related to actual durable knowledge:
- Michelle L. Rivers, Paige E. Northern & Sarah K. Tauber, 20252025DOI: 10.1007/s10648-025-10040-4
- Czyż S. H., Wójcik A. M., Solarská P., Kiper P. — **Year:** 20242024DOI: 10.1038/s41598-024-65753-3
- Wenbo Zhao, Muzi Xu, Chenyuqi Xu, Baike Li, Xiao Hu, Chunliang Yang & Liang Luo, 20232023DOI: 10.3390/jintelligence11100190
- Chunliang Yang, Wenbo Zhao, Bo Yuan, Liang Luo & David R. Shanks, 20222022DOI: 10.3102/00346543221094083
- Jonathan W. Kelly, Alex F. Lim & Shana K. Carpenter, 2021 (issue dated 2022; vol. 11(1):76)2021DOI: 10.1016/j.jarmac.2021.06.001
- Nicholas C. Soderstrom & Robert A. Bjork, 20152015DOI: 10.1177/1745691615569000
- Robert A. Bjork, John Dunlosky & Nate Kornell, 2013 (OpenAlex dates the online-first record 2012)2013DOI: 10.1146/annurev-psych-113011-143823
- Elizabeth Ligon Bjork & Robert A. Bjork, 20112011
- Nate Kornell & Robert A. Bjork, 20082008DOI: 10.1111/j.1467-9280.2008.02127.x
- Bertsch S., Pesta B. J., Wiscott R., McDaniel M. A. — **Year:** 20072007DOI: 10.3758/BF03193441
Interleaving pays off most when the categories being mixed are similar / confusable (Brunmair & Richter 2019) and when the learner is pushed into discriminative contrast — juxtaposing confusable items so their differences become salient (Kang & Pashler 2011). Two /postgres surfaces are almost ideally confusable:
- Andy Tao Li, De Liu, Sean Xin Xu & Cheng Yi, 2024 (MIS Quarterly 48:4, 1363–1394)2024DOI: 10.25300/misq/2023/17206
- Steven C. Pan, Sergio Rodríguez Flores, Michelle E. Kaku & Wing Hei Esmee Lai, 2024 (Learning and Instruction 95:102045)2024DOI: 10.1016/j.learninstruc.2024.102045
- Erdem Onan, Felicitas Biwer, Roman Abel, Wisnu Wiradhany & Anique B. H. de Bruin, 2024 (npj Science of Learning 9)2024DOI: 10.1038/s41539-024-00245-7
- Roman Abel, Anique B. H. de Bruin, Erdem Onan & Julian Roelle, 2024 (Educational Psychology Review)2024DOI: 10.1007/s10648-024-09902-0
- Lea Nemeth & Frank Lipowsky, 2023 (European Journal of Psychology of Education)2023DOI: 10.1007/s10212-023-00723-3
- Joshua Samani & Steven C. Pan, 2021 (npj Science of Learning 6:32)2021DOI: 10.1038/s41539-021-00110-x
- Matthias Brunmair & Tobias Richter, 2019 (Psychological Bulletin)2019DOI: 10.1037/bul0000209
- Rohrer D. — **Year:** 20122012DOI: 10.1007/s10648-012-9201-3
- Sean H. K. Kang & Harold Pashler, 2011 (Applied Cognitive Psychology)2011DOI: 10.1002/acp.1801
- Rohrer D., Taylor K. — **Year:** 20072007DOI: 10.1007/s11251-007-9015-8
Postgres internals is a high prior-knowledge, high-structure domain: B-trees, MVCC, WAL, lock modes, join algorithms. Two evidence streams make it an ideal home for elaboration aids:
- Svetlana Pinet & Marieke Longcamp, 20252025DOI: 10.3389/fpsyg.2024.1517235
- Giuseppe Marano, Georgios D. Kotzalidis, Marianna Mazza, et al., 20252025DOI: 10.3390/life15030345
- F. R. van der Weel & Audrey L. H. van der Meer, 20242024DOI: 10.3389/fpsyg.2023.1219945
- Amber E. Witherby & Shana K. Carpenter, 20212021DOI: 10.1037/xlm0000996
- Ose Askvik E., van der Weel F. R., van der Meer A. L. H. — **Year:** 20202020DOI: 10.3389/fpsyg.2020.01810
- Pam A. Mueller & Daniel M. Oppenheimer, 20142014DOI: 10.1177/0956797614524581
- Marieke Longcamp, Céline Boucard, Jean-Claude Gilhodes, Jean-Luc Anton, Muriel Roth, Bruno Nazarian & Jean-Luc Velay, 20082008DOI: 10.1162/jocn.2008.20504
- Michele M. Dornisch & Rayne A. Sperling, 20062006DOI: 10.3200/joer.99.3.156-166
- Michael Pressley, Mark A. McDaniel, James E. Turnure, Eileen Wood & Maheen Ahmad, 19871987DOI: 10.1037/0278-7393.13.2.291
- Sadoski M. — **Year:** 20052005DOI: 10.1080/10573560590949359
- Sadoski M., Goetz E. T., Rodriguez M. — **Year:** 20002000DOI: 10.1037/0022-0663.92.1.85
- Moreno R., Mayer R. E. — **Year:** 19991999DOI: 10.1037/0022-0663.91.2.358
- Mayer R. E., Anderson R. B. — **Year:** 19911991DOI: 10.1037/0022-0663.83.4.484
- Clark J. M., Paivio A. — **Year:** 19911991DOI: 10.1007/BF01320076
- Paivio A. — **Year:** 19911991DOI: 10.1037/h0084295
- Mayer R. E. — **Year:** 19891989DOI: 10.1037/0022-0663.81.2.240
- Paivio A., Csapo K. — **Year:** 19731973DOI: 10.1016/0010-0285(73)90032-7
Method of loci / memory palace
7 papersThe method of loci excels precisely where Postgres internals are hardest to retain:
- Jingyuan Ren, Boris N. Konrad, Yannan Zhu, Fan Li, Michael Czisch, Martin Dresler, Isabella C. Wagner — 20252025DOI: 10.1101/2025.02.24.639840
- Brigham Moll, Edward R. Sykes — 2022 (Virtual Reality)2022DOI: 10.1007/s10055-022-00700-z
- Isabella C. Wagner, Boris N. Konrad, Philipp Schuster, Sarah Weisig, Dimitris Repantis, Kathrin Ohla, Simone Kühn, Guillén Fernández, Axel Steiger, Claus Lamm, Michael Czisch, Martin Dresler — 2021 (Science Advances)2021DOI: 10.1126/sciadv.abc7606
- Martin Dresler, William R. Shirer, Boris N. Konrad, Nils C. J. Müller, Isabella C. Wagner, Guillén Fernández, Michael Czisch, Michael D. Greicius — 2017 (Neuron 93(5):1227–1235.e6)2017DOI: 10.1016/j.neuron.2017.02.003
- Legge E., Madan C., Ng E., Caplan J. — **Year:** 20122012DOI: 10.1016/j.actpsy.2012.09.002
- Maguire E., Valentine E., Wilding J., Kapur N. — **Year:** 20032003DOI: 10.1038/nn988
- Henry L. Roediger III — 1980 (Journal of Experimental Psychology: Human Learning and Memory 6(5):558)1980DOI: 10.1037/0278-7393.6.5.558
Concrete examples
5 papers- Rawson K. A., Thomas R. C., Jacoby L. L. — **Year:** 20142014DOI: 10.1007/s10648-014-9273-3
- Goldstone R. L., Son J. Y. — **Year:** 20052005DOI: 10.1207/s15327809jls1401_4
- Gentner D., Loewenstein J., Thompson L. — **Year:** 20032003DOI: 10.1037/0022-0663.95.2.393
- Quilici J. L., Mayer R. E. — **Year:** 19961996DOI: 10.1037/0022-0663.88.1.144
- Gick M. L., Holyoak K. J. — **Year:** 19831983DOI: 10.1016/0010-0285(83)90002-6
Contextual interference
5 papers- Brady F. — **Year:** 20042004DOI: 10.2466/pms.99.1.116-126
- Brady F. — **Year:** 19981998DOI: 10.1080/00336297.1998.10484285
- Magill R. A., Hall K. G. — **Year:** 19901990DOI: 10.1016/0167-9457(90)90005-X
- Goode S. L., Magill R. A. — **Year:** 19861986DOI: 10.1080/02701367.1986.10608091
- Shea J. B., Morgan R. L. — **Year:** 19791979DOI: 10.1037/0278-7393.5.2.179
- Dunlosky J., Rawson K.A., Marsh E.J., Nathan M.J., Willingham D.T. — **Year:** 20132013DOI: 10.1177/1529100612453266
- Bjork R.A., Dunlosky J., Kornell N. — **Year:** 20132013DOI: 10.1146/annurev-psych-113011-143823
- Metcalfe J. — **Year:** 20092009DOI: 10.1111/j.1467-8721.2009.01628.x
- Metcalfe J., Kornell N. — **Year:** 20052005DOI: 10.1016/j.jml.2004.12.001
- Nelson T.O., Narens L. — **Year:** 19901990DOI: 10.1016/S0079-7421(08)60053-5
Self-explanation
5 papers- Bisra K., Liu Q., Nesbit J. C., Salimi F., Winne P. H. — **Year:** 20182018DOI: 10.1007/s10648-018-9434-x
- Rittle-Johnson B., Loehr A. M., Durkin K. — **Year:** 20172017DOI: 10.1007/s11858-017-0834-z
- Rittle-Johnson B. — **Year:** 20062006DOI: 10.1111/j.1467-8624.2006.00852.x
- Chi M. T. H., De Leeuw N., Chiu M.-H., Lavancher C. — **Year:** 19941994DOI: 10.1207/s15516709cog1803_3
- Chi M. T. H., Bassok M., Lewis M. W., Reimann P., Glaser R. — **Year:** 19891989DOI: 10.1207/s15516709cog1302_1
Spaced-repetition algorithms
5 papers- Su J., Ye J., Nie L., Cao Y., Chen Y. — **Year:** 20232023DOI: 10.1109/TKDE.2023.3251721
- Ye J., Su J., Cao Y. — **Year:** 20222022DOI: 10.1145/3534678.3539081
- Lindsey R. V., Shroyer J. D., Pashler H., Mozer M. C. — **Year:** 20142014DOI: 10.1177/0956797613504302
- Pavlik P. I., Anderson J. R. — **Year:** 20082008DOI: 10.1037/1076-898X.14.2.101
- Pavlik P. I., Anderson J. R. — **Year:** 20052005DOI: 10.1207/s15516709cog0000_14
Worked examples
5 papers- Kalyuga S., Ayres P., Chandler P., Sweller J. — **Year:** 20032003DOI: 10.1207/s15326985ep3801_4
- Atkinson R. K., Derry S. J., Renkl A., Wortham D. — **Year:** 20002000DOI: 10.3102/00346543070002181
- Renkl A. — **Year:** 19971997DOI: 10.1207/s15516709cog2101_1
- Paas F. G. W. C., Van Merriënboer J. J. G. — **Year:** 19941994DOI: 10.1037/0022-0663.86.1.122
- Sweller J., Cooper G. A. — **Year:** 19851985DOI: 10.1207/s1532690xci0201_3
Chunking
4 papers- Gobet F., Lane P. C. R., Croker S., Cheng P. C-H., Jones G., Oliver I., Pine J. M. — **Year:** 20012001DOI: 10.1016/S1364-6613(00)01662-4
- Cowan N. — **Year:** 20012001DOI: 10.1017/S0140525X01003922
- Chase W. G., Simon H. A. — **Year:** 19731973DOI: 10.1016/0010-0285(73)90004-2
- Miller G. A. — **Year:** 19561956DOI: 10.1037/h0043158
Concept mapping
4 papers- Karpicke J. D., Blunt J. R. — **Year:** 20112011DOI: 10.1126/science.1199327
- Nesbit J. C., Adesope O. O. — **Year:** 20062006DOI: 10.3102/00346543076003413
- Novak J. D. — **Year:** 20022002DOI: 10.1002/sce.10032
- Heinze-Fry J. A., Novak J. D. — **Year:** 19901990DOI: 10.1002/sce.3730740406
Deliberate practice
4 papers- Macnamara B. N., Hambrick D. Z., Oswald F. L. — **Year:** 20142014DOI: 10.1177/0956797614535810
- Hambrick D. Z., Oswald F. L., Altmann E. M., Meinz E. J., Gobet F., Campitelli G. — **Year:** 20142014DOI: 10.1016/j.intell.2013.04.001
- Ericsson K. A. — **Year:** 20082008DOI: 10.1111/j.1553-2712.2008.00227.x
- Ericsson K. A., Krampe R. T., Tesch-Römer C. — **Year:** 19931993DOI: 10.1037/0033-295X.100.3.363
Elaborative interrogation
4 papers- Smith B.L., Holliday W.G., Austin H.W. — **Year:** 20102010DOI: 10.1002/tea.20378
- Woloshyn V.E., Wood E., Willoughby T. — **Year:** 19941994DOI: 10.1002/acp.2350080104
- Woloshyn V.E., Willoughby T., Wood E., Pressley M. — **Year:** 19901990DOI: 10.1037/0022-0663.82.3.513
- Pressley M., McDaniel M.A., Turnure J.E., Wood E., Ahmad M. — **Year:** 19871987DOI: 10.1037/0278-7393.13.2.291
Feedback timing
4 papers- Van der Kleij F. M., Feskens R. C. W., Eggen T. J. H. M. — **Year:** 20152015DOI: 10.3102/0034654314564881
- Shute V. J. — **Year:** 20082008DOI: 10.3102/0034654307313795
- Hattie J., Timperley H. — **Year:** 20072007DOI: 10.3102/003465430298487
- Kulik J. A., Kulik C. C. — **Year:** 19881988DOI: 10.3102/00346543058001079
Note-taking strategies
4 papers- Kobayashi K. — **Year:** 20062006DOI: 10.1111/j.1468-5884.2006.00311.x
- Kobayashi K. — **Year:** 20052005DOI: 10.1016/j.cedpsych.2004.10.001
- Kiewra K. A. — **Year:** 19891989DOI: 10.1007/BF01326640
- Di Vesta F. J., Gray G. S. — **Year:** 19721972DOI: 10.1037/h0032243
Pretesting effect
4 papers- Little J. L., Bjork E. L. — **Year:** 20162016DOI: 10.3758/s13421-016-0621-z
- Potts R., Shanks D. R. — **Year:** 20142014DOI: 10.1037/a0033194
- Richland L. E., Kornell N., Kao L. S. — **Year:** 20092009DOI: 10.1037/a0016496
- Kornell N., Hays M. J., Bjork R. A. — **Year:** 20092009DOI: 10.1037/a0015729
Generation effect
3 papers- DeWinstanley P. A., Bjork E. L. — **Year:** 20042004DOI: 10.3758/BF03196872
- Crutcher R. J., Healy A. F. — **Year:** 19891989DOI: 10.1037/0278-7393.15.4.669
- Slamecka N. J., Graf P. — **Year:** 19781978DOI: 10.1037/0278-7393.4.6.592
The principles behind the techniques
The catalogue above lists the techniques and the papers behind them. This section explains why they work: thirteen principles across encoding (what makes a trace strong at the moment it forms), practice (how strength grows or decays), and structure (the formats that organize material) — each with its boundary conditions and the concrete constraints it implies for generating memory cues at scale. Where recent work (2023–2026) has built a system that puts a principle to work, an In generation systems note records it, each finding traced to a verified source. Several principles also carry a Machine-learning analog — the same idea rediscovered as a training mechanism (replay, contrastive learning, retrieval augmentation, curriculum learning).
Encoding principles
What happens at the moment a memory is formed. These seven principles determine the starting strength of a trace — how vivid, distinctive and well-bound the association is when it first lands. For a mnemonic generator, they are the ranking and generation constraints.
Imageability and dual coding
Paivio’s dual coding theory holds that concrete, picturable content gets encoded in two independent systems — verbal and imaginal — giving it two retrieval routes instead of one. The effect size is large and among the most replicated in memory research.
Imageability is also measurable: psycholinguistic norms (the MRC database, Brysbaert’s concreteness norms covering ~40k English words) let you score generated cues automatically, and newer work scores sentence-level imageability by feeding text to a text-to-image model and measuring generation quality and consistency.
- Concrete words are recalled roughly twice as well as abstract words in paired-associate tasks, and rated imageability predicts recall better than frequency, familiarity, or word length. [Paivio (1991)]
- Concreteness norms exist for ~40,000 English lemmas, making automatic imageability scoring practical at generation time. [Brysbaert et al. (2014)]
The subtlety for a generator: imageability is a property of the whole cue, not the words in it. “Justice crushed the melon” is more imageable than its abstract subject suggests because the scene supplies the image.
- Score at the sentence level, not by averaging word norms — word-level norms will systematically underrate good mnemonics for abstract vocabulary, which is exactly where users need help most.
- Sentence-level imageability is now directly measurable: feed the cue to a text-to-image model and score the generations for quality and cross-sample consistency — the operational form of “score the whole cue, not the words in it.” [Wu & Smith (2023)]
- Vision–language models track human concreteness ratings closely enough to serve as automatic concreteness scorers, extending norm-based screening past the ~40k words the published norms cover — exactly the abstract-vocabulary regime where generated cues need the check most. [Roy et al. (2026)]
Multimodal joint embeddings (CLIP). The machine-learning analog is multimodal representation learning: CLIP encodes the same content through two parallel channels — an image encoder and a text encoder — and aligns them in one shared space, the engineering echo of storing a concept as mutually reinforcing visual and verbal codes. [Radford et al. (2021)]
The bizarreness effect
Bizarreness operates through two mechanisms: distinctiveness at encoding (unusual items get more elaborative processing) and reduced output interference at retrieval (bizarre items form a smaller, more discriminable retrieval set). Both mechanisms are relative — they depend on bizarre items being rare in their context.
- The canonical result: bizarre sentences (“The dog rode the bicycle down the street”) are recalled better than common ones (“The dog chased the bicycle down the street”), but almost exclusively in mixed-list designs. When every item is bizarre, the advantage shrinks to near zero and sometimes reverses. [McDaniel & Einstein (1986)]
Bizarreness helps item access (did you remember the sentence at all?) but can slightly hurt associative binding(which word went with which cue?) if the bizarre elaboration is disconnected from the target. Weird-for-weird’s-sake decoration is worse than a plausible cue tightly bound to the target.
The effect is strongest at free recall and weaker at cued recall — and keyword mnemonics are a cued-recall task. So bizarreness is the seasoning, not the foundation: the binding (principle 5) does the heavy lifting; bizarreness buys distinctiveness on top.
- Intensity should vary within a study set — a batch scorer that enforces a distribution of vividness across cards will outperform per-card maximization.
- The bizarre element must involve the keyword-meaning link itself, not be bolted on.
- Whether the mixed-list moderation holds for AI-generated mnemonics at app scale is untested — a genuinely publishable eval.
- Overgenerate-and-rank pipelines give the principle a concrete home: a keyword-mnemonic generator proposes many candidate cues and a scoring stage ranks them, and that stage is exactly where a rarity or diversity penalty enforces the mixed-list distribution — vary vividness across a study set rather than maximizing bizarreness on every card. [Lee et al. (2024)]
- “Rare in its context” is becoming measurable rather than guessed: representational distinctiveness within a category predicts memorability, so a generator can score how much a cue stands out from its neighbors instead of assuming weirder is always better. [Atzert et al. (2026)]
Prioritized experience replay. Reinforcement learning independently rediscovered the effect: prioritized experience replay samples stored transitions in proportion to their temporal-difference error, so the most surprising experiences — the ones the model predicted worst — are replayed and consolidated disproportionately often. [Schaul et al. (2015)]
Distinctiveness (von Restorff / isolation effect)
The 1933 von Restorff finding: an item that differs from its surroundings — in color, size, category, anything — is recalled dramatically better. Modern accounts (Hunt’s item-specific vs. relational processing framework) generalize it: memory benefits from processing what makes an item different from its neighbors, and separately from processing what items share.
Best recall comes from doing both — organizing items relationally (these are all Spanish food words) while encoding each distinctively (this one has an exploding tomato).
- The original isolation-effect demonstration: any dimension of difference from surrounding items produces a large recall advantage. [von Restorff (1933)]
- Item-specific and relational processing contribute separately; combining them beats either alone. [Hunt & McDaniel (1993)]
Distinctiveness is relative to the set. LLMs left to their own devices converge on stock imagery — everything becomes an elephant or an explosion — which silently destroys this benefit even when every card scores well individually.
- Set-aware generation: when a user studies 20 words, the 20 cues should be distinct from each other in scene, setting, and imagery.
- A diversity penalty across a batch is cheap to implement and directly targets the mechanism.
- Overgenerate-and-rank pipelines already exist for keyword mnemonics — sample many candidate cues per item, then rank on psycholinguistic measures. The ranking stage is the natural insertion point for a set-aware diversity penalty, which per-candidate scoring alone cannot supply. [Lee et al. (2024)]
- Distinctiveness is becoming directly measurable: scene memorability tracks how distinct an item’s representation is within its own visual category, and caption-based frameworks tie image memorability to linguistic features — so “distinct within its set” is a property a batch scorer can compute rather than guess. [Atzert et al. (2026)][Koneru & Powell (2026)]
Contrastive hard-negative mining. Contrastive learning does the same job in embedding space: hard-negative mining finds the most confusable items and pushes their representations maximally apart, so look-alikes become separable — distinctiveness enforced by the training objective. [Robinson et al. (2020)]
Elaboration and levels of processing
Craik and Lockhart’s framework: memory is a byproduct of processing depth. Semantic processing (what does it mean, how does it relate) beats phonemic (what does it sound like) beats orthographic (what does it look like). Craik and Tulving’s follow-up added the congruityrefinement: elaboration helps most when it forms an integrated, coherent structure around the target — richness alone isn’t enough; the elaboration has to make the target more predictable from the cue.
- Semantic orienting tasks produce substantially better retention than phonemic or orthographic ones, holding exposure constant. [Craik & Lockhart (1972)]
- Elaboration helps most when the elaborated context is congruous with the target — an integrated structure, not decoration. [Craik & Tulving (1975)]
Keyword mnemonics are interesting here because they deliberately start shallow — the keyword link is phonemic (“carta” sounds like “cart”) — and then rescue it with a semantic layer (the verbal cue: “a letter riding in a shopping cart”). The evidence says the semantic layer is where the value lives.
- Practical test for a generated cue: could a user who remembers the scene derivethe meaning from it? If the scene is memorable but the meaning isn’t recoverable from it, the cue fails congruity.
- That’s a checkable property — prompt a second model with only the scene and ask it to guess the definition; use recovery rate as a ranking signal.
- Elaboration quality can be learned from learners: mixing an LLM mnemonic generator with student feedback on which mnemonics actually stick aligns generation with learning outcomes rather than surface cleverness — an empirical stand-in for the congruity test, since cues whose elaboration integrates the meaning are the ones students keep. [Balepur et al. (2024)]
Masked / self-supervised pretraining. Levels-of-processing shows up in how models are pretrained: BERT’s masked-language-model objective can’t be solved with surface co-occurrence — recovering a masked token from full context forces deep semantic encoding, the same reason deep processing yields more durable traces than shallow. [Devlin et al. (2019)]
Interactive imagery
In Atkinson and Raugh’s original keyword studies (1975, Russian vocabulary — the paradigm every keyword mnemonic descends from), the instruction that produced the large gains was specifically to form an image in which keyword and referent interact in a single scene. Motion verbs almost force interaction — things colliding, chasing, crushing, balancing on each other — which is the strongest scientific justification for a “motion encouraged” style rule.
- The keyword method with interactive-imagery instructions produced large vocabulary-learning gains over rote controls. [Atkinson & Raugh (1975)]
- Interacting images beat separated images substantially — and, notably, interaction mattered more than bizarreness. A picture of a piano smoking a cigar beats a bizarre-but-separate piano and cigar. [Wollen, Weber & Lowry (1972)]
Interaction is the load-bearing variable; bizarreness (principle 2) only adds distinctiveness on top of an already-interacting scene.
- A concrete generation constraint: the keyword entity and the meaning entity should share a verb. “The cart carries the letter” passes; “there is a cart, and also a letter” fails.
- This is trivially checkable with dependency parsing or a cheap LLM judge — probably the highest-value single rule a generator can enforce.
- The full Atkinson-Raugh pipeline is now automated end to end: an LLM selects the keyword and writes the verbal cue, and a text-to-image model renders the scene — which puts the interaction constraint exactly where it’s enforceable, on the verbal cue before the image is generated. [Lee & Lan (2023)]
Relational binding (Relation Networks). Interaction-in-one-scene has a representation-space twin: a Relation Network takes each pair of entities and fuses them through a shared network into one joint relational embedding rather than storing them side by side — binding interacting elements exactly as interactive imagery does. [Santoro et al. (2017)]
The generation effect
Material you produce yourself, even trivially (completing “hot–c___” vs. reading “hot–cold”), is remembered better than material you receive. Effect sizes are moderate but robust, and the effect extends to self-generated mnemonics vs. provided ones.
- The original delineation of the phenomenon across five experiments, with even minimal generation producing reliable gains. [Slamecka & Graf (1978)]
The important twist: provided mnemonics from a good source often beat self-generated ones from novices, because novices generate low-quality cues. Quality and ownership trade off.
- The design synthesis the literature supports: provide a high-quality generated cue, then get the user to do something generative with it — edit one word, pick between two variants, visualize it for three seconds and rate its vividness. Even minimal generative involvement recovers much of the effect.
- A “select a tamer style” control is incidentally a weak form of this; a per-card “remix” button would be a stronger one.
- The quality–ownership trade-off is being engineered directly: multimodal LLM scaffolding walks the learner through creating their own keyword mnemonic instead of handing one over — the system supplies the quality floor, the learner supplies the generation. [Shao et al. (2026)]
Self-generated training data (STaR). The model-side echo of self-produced material: STaR fine-tunes a model on the correct chain-of-thought rationales it generated itself, so its own produced reasoning becomes its training signal — the generation effect turned into a bootstrap loop. [Zelikman et al. (2022)]
Emotional arousal and humor
Arousing content gets consolidation priority — the amygdala modulates hippocampal encoding, and the advantage for emotional material actually grows over a delay. This matters for a learning product because the outcome of interest is delayed retention, not immediate recall. The workplace-safe channels into arousal are humor and surprise.
- The classic finding: arousing pairs were recalled worse than neutral at immediate test but much better a week later. [Kleinsmith & Kaplan (1963)]
- Humorous sentences are recalled better than matched non-humorous ones — with the same mixed-list caveat as bizarreness. [Schmidt (1994)]
Humor draws attention to itself: if the joke is about something other than the keyword-meaning link, it can cannibalize memory for neighboring content and for the link itself. Same congruity rule as principle 4 — the funny thing must be the association.
There is no good evidence that sexual or gory content outperforms humorous/surprising content for verbal-associative learning at delayed test. Taboo words show attentional-capture effects, but capture isn’t binding — a content policy that excludes them gives up little or nothing, which makes the policy scientifically defensible, not just commercially prudent.
- Use humor and surprise as the arousal channels; keep the funny element on the keyword-meaning link itself.
- Optimize for delayed retention, not immediate-recall metrics — the arousal advantage only shows up after consolidation.
- The overgenerate-and-rank pattern extends to arousal: a generator proposes many candidates and scores each for humor or surprise as one ranking signal, keeping the funny element on the keyword-meaning link rather than maximizing raw novelty per card. [Lee et al. (2024)]
- Because the arousal advantage only appears at delayed test, it can’t be judged at generation time — so systems close the loop with learner feedback on which cues actually stick, aligning the generator to retention rather than to a generation-time humor score. [Balepur et al. (2024)]
Salience-weighted replay. Machine learning collapses what human memory splits: prioritized experience replay weights consolidation by TD-error, a single priority signal that stands in for both surprise (bizarreness) and value/salience (arousal) — the same replay mechanism cited under bizarreness, viewed through the salience lens. [Schaul et al. (2015)]
Practice principles
How trace strength grows — or decays — after encoding. These are the largest effects in all of applied psychology, and they belong to the review schedule, not the cue. A system that nails encoding and ignores practice produces impressive demos and poor month-three retention.
The spacing effect
From Ebbinghaus (1885) through Cepeda et al.’s 2006 meta-analysis (839 assessments): distributed practice beats massed practice, everywhere, at every age, for every material type tested. Expanding schedules (1 day, 3 days, 10 days, 30 days) are what SuperMemo/Anki-style algorithms implement, and the modern end of this line (FSRS, Duolingo’s half-life regression) fits individual forgetting curves per item per user.
- The forgetting curve and the first demonstration that distributed repetition flattens it. [Ebbinghaus (1885)]
- 839 assessments of distributed practice: the effect holds across ages, materials and retention intervals. [Cepeda et al. (2006)]
- The optimal gap scales with the retention interval — roughly 10–20% of the time until you need the memory. [Cepeda et al. (2008)]
- Individual forgetting curves can be fit per item per user at product scale. [Settles & Meeder (2016)]
The blunt product truth: scheduling dominates cue quality. A d≈0.4–0.6 improvement from better mnemonics is real money, but spacing vs. massing is one of the largest effects in all of applied psychology.
- If a product doesn’t own the review schedule, it should integrate with something that does (Anki export is the obvious move). This site’s review queue implements FSRS-6 for exactly this reason.
- The scheduler lineage is now data-driven end to end: half-life regression fits a forgetting curve per item per user, and the open FSRS family that succeeds it powers most modern spaced-repetition apps — the spacing effect turned into a trained model rather than a fixed expanding ladder. [Settles & Meeder (2016)]
- That lineage keeps advancing: deep-reinforcement-learning schedulers now estimate recall probability with a Transformer and optimize when to review, pushing the scheduler past fixed-form regression fits toward a policy learned from review histories (computational model). [Xiao & Wang (2024)]
- The effect is even being reproduced inside generative systems: a computational model casts hippocampo-neocortical consolidation as compressive retrieval-augmented generation, replaying compressed memories to train a generative network — spacing and consolidation implemented in silico (model-level analogy; preprint). [Spens & Burgess (2024)]
Experience replay (continual learning). Neural networks forget too, and the fix is the same shape: replay methods re-expose the network to buffered past examples interleaved with new learning, spacing old material back in to fight catastrophic forgetting — spaced re-exposure as an anti-forgetting algorithm. [Rolnick et al. (2019)]
Retrieval practice (the testing effect)
Mechanistically, successful effortful retrieval strengthens and multiplies retrieval routes in a way that re-exposure doesn’t. The catch is metacognitive: restudy feels more effective than testing, so users will ask for reread modes that hurt them — an illusion the UX has to fight.
- After a week, students who practiced retrieving a passage retained ~50% more than students who restudied it — even though restudy felt more effective to the students themselves. [Roediger & Karpicke (2006)]
- Retrieval practice even beats elaborative concept mapping. [Karpicke & Blunt (2011)]
For mnemonics specifically: the goal state is that the learner eventually retrieves the meaning directly and the mnemonic scaffolding drops away — which is what the longitudinal keyword-method studies show happens naturally with practice.
- Test in the direction of eventual use (see target word → recall meaning).
- Have the mnemonic serve as a fallback scaffold — shown only after an attempt, or on failure.
- Retrieval practice is only as good as its questions, and automatic question generation now supplies them at scale — recent systems generate assessment items with independent control over topic and difficulty, which is what lets a tutor test in the direction of use and tune the retrieval effort per item rather than reusing a fixed bank. [Eldho Paul & Sunar (2026)]
Associative memory / pattern completion. Retrieval-as-reconstruction is the mechanism of associative memory: a modern Hopfield layer takes a partial or noisy cue and completes it to the stored pattern in a single update — the computational form of producing the answer from a cue rather than re-reading it. [Ramsauer et al. (2020)]
Desirable difficulties
Bjork’s umbrella framework unifying spacing, testing, interleaving, and generation: conditions that slow apparent learning often enhance long-term retention, because retrieval strength and storage strength are distinct, and effortful processing builds the latter.
Two additions to what’s covered elsewhere on this page: interleaving (mixing categories — ABCABC — beats blocking — AAABBB — for discrimination learning; relevant when users learn confusable sets like ser/estar or kanji with shared radicals) and the retrieval-effort gradient (harder successful retrievals strengthen more — the theoretical basis for scheduling reviews just before predicted forgetting).
- The framework distinguishing retrieval strength from storage strength, and the case that effortful conditions build durable learning. [Bjork (1994)]
The caveat Bjork himself stresses: difficulties are desirable only when the learner can overcome them. For beginners, too much difficulty means failed retrievals and frustration.
- Adaptive difficulty: richer mnemonic support early, progressively stripped as items mature.
- Interleave confusable material rather than blocking it — the review queue should never show two of the same confusable category in a row.
- The “desirable” in desirable difficulty is now a tunable knob: multi-agent frameworks generate items at a targeted difficulty under feature constraints, so a system can hold each learner near the edge of what they can just overcome rather than guessing. [Hwang et al. (2026)]
- The support side is being personalized too: LLM systems adapt the generated mnemonic to each learner’s preferences, which is the surface that lets scaffolding be rich early and progressively stripped as items mature. [Lee et al. (2026)]
- And the target itself is being measured, not guessed: open-ended language tutors now align difficulty to a learner’s trajectory, steering toward the optimal-challenge zone — the operational form of Bjork’s caveat that difficulty helps only when it can be overcome. [Shu et al. (2026)]
Curriculum learning. Curriculum learning is the training-time instantiation: ordering examples meaningfully from easy to hard speeds convergence and improves generalization — difficulty treated as a schedulable variable, desirable exactly when the model can keep up. [Bengio et al. (2009)]
Structural principles
Formats that organize material before and around encoding: recoding raw items into chunks, anchoring them to space, and keeping the encoding-retrieval match intact. The last of these — encoding specificity — is the invariant that keeps the whole loop connected.
Chunking
Miller (1956) and the modern revision (Cowan: ~4 items, not 7): working memory is limited in chunks, not bits, so recoding raw material into meaningful units multiplies capacity. This is the mechanism behind all expert memory.
- The original span limit, framed in chunks. [Miller (1956)]
- The modern re-estimate: ~4 chunks of genuinely independent capacity. [Cowan (2001)]
- A runner extended digit span from 7 to ~80 by chunking digits into running times. [Ericsson et al. (1980)]
Chunking matters most beyond single vocabulary items: acronym mnemonics, number-shape/number-rhyme systems, and the Major system (digits→consonants→words) are all chunking schemes.
- Chunking composes with everything above — chunk first, then make the chunk vivid.
- Chunking is now an explicit generation step: a system decomposes each kanji into its components and generates a mnemonic per component with an EM algorithm — chunk first, then make each chunk vivid, exactly as the principle prescribes. [Lee et al. (2025)]
- It also emerges rather than being hand-coded: in an RL-optimized model of Turkish inflection, memory-retrieval pressure alone drives the recoding of morpheme sequences into reusable multi-affix chunks — the capacity-multiplying move appearing as an optimization (computational model). [Elsner et al. (2026)]
Subword tokenization (BPE). Every language model chunks its input first: byte-pair encoding iteratively merges the most frequent adjacent symbols into single reusable subword tokens, recoding a stream of characters into meaningful units — chunking as the very first layer of the stack. [Sennrich et al. (2016)]
Method of loci
The oldest technique in the record (Simonides, ~500 BCE) and still the most powerful for orderedmaterial. It works by parasitizing spatial memory, which is evolutionarily old, high-capacity, and comes with built-in ordering and built-in cues (walk the route again). It is also inherently a scene-action-motion format — a person doing something absurd at a location — so a “memory palace mode” sits entirely inside a motion-encouraged style policy.
- Memory champions have ordinary memory capacity but reorganized, spatially-anchored retrieval strategies. [Maguire et al. (2002)]
- Six weeks of loci training in novices durably improved recall and shifted their brain connectivity toward the memory-athlete pattern. [Dresler et al. (2017)]
Loci content must be user-anchored (their apartment, their commute) — a generic palace loses the built-in cues that make the technique work.
- Elicit a route once, then generate scene-per-locus — a personalization surface none of the recent mnemonic-generation papers touch.
- The scene-per-locus idea is now a built system: a generated memory palace in VR places an object at each locus and writes an association per pair, with the virtual route supplying the ordering — benchmarked against random pairing at immediate and one-week recall, so the spatial scaffold is evaluated, not just asserted. [Wulff et al. (2026)]
Addressable external memory (DNC / NTM). The method of loci is external addressable memory: a Differentiable Neural Computer writes items to memory locations and keeps a temporal-link matrix of write order, so it can traverse memory in sequence — the machine version of storing items along a route and walking it back. [Graves et al. (2016)]
Encoding specificity and transfer-appropriate processing
Tulving and Thomson (1973): a cue works at retrieval only to the extent it was part of the encoding. The famous demonstrations are context effects (divers recalling underwater-learned words better underwater), but the operational version for a mnemonic system is stricter and more mundane: the exact cue shown at review must match the cue used at encoding.
The companion principle, transfer-appropriate processing, says match practice to the criterion task: if the user needs to produce the word when speaking, practice production, not just recognition — multiple-choice review builds recognition that doesn’t fully transfer to recall.
- The encoding-specificity principle: retrieval succeeds when the cue was encoded with the trace, and fails when it wasn’t — even for strong semantic associates. [Tulving & Thomson (1973)]
- Divers recalled underwater-learned words better underwater — the classic context-dependence demonstration. [Godden & Baddeley (1975)]
If a system regenerates mnemonics between sessions, or A/B-tests cue variants on the same item for the same user, it severs the encoding-retrieval match and the mnemonic silently stops working — the user re-encodes from scratch while the metrics say they’re reviewing.
- Cue immutability per (user, item) should be a hard invariant in the data model, with regeneration treated as a reset event.
- Match review format to the criterion task — production practice for production goals, not multiple choice.
- Cross-lingual keyword generators now build the cue from the encoding overlap itself: a system retrieves an L1 keyword sequence by IPA-level phonological alignment to the L2 word, so the sound that bridges the two at study is exactly the cue present at recall — encoding specificity engineered into the generator. [Kang et al. (2025)]
Retrieval-augmented generation. RAG is encoding specificity in vector space: the query is encoded into a cue that must land near the passage embeddings written into the store, so a memory is retrieved only when the retrieval cue’s encoding overlaps the one used to write it — cue must match trace. [Lewis et al. (2020)]
How it composes
The dependency structure matters more than any single principle. Encoding quality (principles 1–7) determines the starting strength of a trace; retrieval practice (9–10) determines how that strength grows; spacing (8) determines whether growth compounds or decays; and specificity (13) is the invariant that keeps the whole loop connected. A system that nails 1–7 and ignores 8–9 produces impressive demos and poor month-three retention.
The keyword-method literature learned this the hard way — early studies showed big immediate gains that sometimes faded at long delays when there was no spaced retrieval, which critics wrongly read as “mnemonics don’t last” when the actual lesson was “encoding techniques need a practice schedule.”
One last pattern is worth naming: every one of these principles was independently rediscoveredby machine learning as a training mechanism (see each card’s Machine-learning analog). Spacing is experience replay against catastrophic forgetting; distinctiveness is contrastive hard-negative mining; encoding specificity is retrieval-augmented generation; the generation effect is training on self-produced rationales; method of loci is addressable external memory. The convergence is telling — these aren’t arbitrary study tricks but general constraints on any system, biological or artificial, that has to store and retrieve under interference. Tellingly, ML collapses some distinctions humans draw: prioritized replay is the single mechanism behind both bizarreness (surprise) and arousal (salience).
- 1Interactive imagery
The interaction constraint on generation — keyword and meaning sharing a verb — is a ranking-stage feature shippable in a week, and probably the highest-value single rule.
- 2Elaboration and levels of processing
The meaning-recoverability check (can a second model derive the definition from the scene alone?) directly tests congruity — the property that separates memorable-and-useful from merely memorable.
- 3Distinctiveness (von Restorff / isolation effect)
Batch-level diversity and intensity variation (with principle 2) protect the relative effects that per-card scoring silently destroys.
- 4Encoding specificity and transfer-appropriate processing
Cue immutability per (user, item) is a data-model invariant — cheap to enforce now, expensive to retrofit after metrics have been corrupted by silent re-encoding.
- 5The spacing effect
Spaced retrieval integration (with principle 9) is the one that determines whether the product actually works at month three — the largest effect, and the last because it’s an integration, not a ranking feature.
How LlamaIndex reflects each principle
Each principle above is not just documented on this page — it is running code in the LlamaIndex pipelines that build this site. The cards below pair every principle with the concrete mechanism that embodies it: the API surface that implements it, the repo file it lives in, and how the grounding pipeline reflects the principle in practice.
The QueryFusionRetriever merges two retrievers: a dense search from embeddings and a BM25 keyword retriever, then combines their results using reciprocal rank fusion. It also uses an LLM to generate additional queries for better coverage. Without it, the system would retrieve only from one method, missing relevant code or prose that the other method finds. It is built in _build_retriever and returns the fused retriever used for context gathering.
Paivio's dual coding theory states concrete, picturable content is encoded in two independent systems—verbal and imaginal—giving it two retrieval routes instead of one. However, imageability is a property of the whole cue, not its individual words, so a scene like "Justice crushed the melon" can be highly imageable despite abstract components. Hybrid retrieval (dense + keyword fusion) applies this by offering two independent retrieval paths, analogous to dual coding, allowing the system to locate a concrete cue through either its semantic or lexical representation.
Mirrors Hybrid retrieval (dense + keyword fusion) in the LlamaIndex primer.
The BM25Retriever is built inside build_fusion_retriever from the provided nodes and a similarity_top_k value. It returns a sparse retriever that scores chunks by lexical term overlap. Its outputs feed into QueryFusionRetriever alongside a dense retriever, merging both rankings via reciprocal rank fusion. Without it, hybrid recall would fall back to the dense retriever alone, losing coverage of exact token matches like identifiers or CLI flags.
The bizarreness effect operates through distinctiveness at encoding—unusual items receive more elaborative processing—and reduced output interference at retrieval due to a smaller, more discriminable retrieval set, but both mechanisms rely on bizarreness being rare in context. Boundaries: it aids free recall but can hurt associative binding if the bizarre element is disconnected, and the effect weakens under cued recall. A Keyword (BM25) exact-match retrieval pipeline stage applies this by weighting rare terms more heavily, mirroring how distinctiveness reduces interference and surfaces unusual cues.
Mirrors Keyword (BM25) exact-match retrieval in the LlamaIndex primer.
FastEmbedRerank is a node-postprocessor that takes a list of NodeWithScore objects and a QueryBundle. It uses FastEmbed’s cross-encoder to score each (query, chunk) pair, squashes the unbounded logits via sigmoid to a [0,1] scale, then sorts descending and deduplicates near-duplicates before returning the top_n. Without this reranker, the bi-encoder’s cosine‑similarity ordering would remain, which often buries truly relevant chunks under near‑duplicate candidates.
The von Restorff isolation effect demonstrates that an item differing from its neighbors in any dimension is recalled dramatically better, with modern accounts showing that memory benefits from processing both what makes an item unique (item-specific) and what items share (relational), and combining both yields the best recall. Boundary conditions from the source state that distinctiveness is relative to the set, and when large language models generate stock imagery, everything becomes similar, silently destroying this benefit. In an overgenerate-and-rank pipeline, the ranking stage—which a Cross-encoder reranker exemplifies—is the natural insertion point for a set-aware diversity penalty that enforces distinctiveness among cues, directly applying the mechanism to prevent convergence on uniform imagery.
Mirrors Cross-encoder reranking in the LlamaIndex primer.
The function build_index_from_nodes creates a VectorStoreIndex over hand-built TextNodes, syncing each batch under a content-keyed synthetic ref_doc_id via sync_nodes. Its inputs are nodes, an embed model, and a namespace; it returns an index backed by a stable collection. Unchanged batches produce zero writes; changed batches land under a new ID (append-only). Without it, there would be no mechanism to idempotently update the vector store or scope retrieval to the latest batch.
Memory is a byproduct of processing depth: semantic processing retains better than phonemic or orthographic. Elaboration helps most when it forms an integrated, coherent structure where the target is predictable from the cue—richness alone isn’t enough. An index that stores cues with recoverable semantic content directly applies this mechanism by ensuring each indexed representation embeds meaning that can be derived from the stored cue.
Mirrors The index in the LlamaIndex primer.
The function make_llm creates a LlamaIndex chat model (a DeepSeekLLM or OpenAILike subclass) by first checking for a DEEPSEEK_API_KEY to use DeepSeek directly, then an LLM_BASE_URL for a local OpenAI-compatible server, and finally falling back to the Cloudflare AI Gateway. It accepts temperature and optional max_tokens as inputs, returning the configured LLM instance. Without it, LlamaIndex components would rely on OpenAI defaults, requiring an OPENAI_API_KEY that the system avoids.
Interactive imagery requires forming a mental image where the keyword and its meaning interact within a single scene; motion verbs that force collision or chasing are the strongest mechanism. Its boundary condition is that interaction itself is the load-bearing variable, with bizarreness only adding distinctiveness on top of an already-interacting scene. The LLM behind the engine puts this to work by generating verbal cues that enforce a shared verb between keyword and meaning, making the interaction constraint trivially checkable before any image is rendered.
Mirrors The LLM behind the engine in the LlamaIndex primer.
The function _ground_code_excerpt retrieves a real code excerpt from a specified file by querying an LLM with a prompt, then verifies every identifier in the returned code belongs to the source's whole-token set. Without this verification, hallucinated APIs would be placed into the page as authoritative examples. The function loops with retries, and if the excerpt still contains violations after all attempts, it returns an empty string to prevent ungrounded code from appearing.
The generation effect is the principle that material you produce yourself, even as trivial as completing "hot–c___" rather than reading "hot–cold," is remembered better than material you receive. However, a boundary condition exists: provided mnemonics from a good source often beat self-generated ones from novices because novices produce low-quality cues, creating a trade-off between quality and ownership. Applying this mechanism, the pipeline stage "From index to page" could prompt users to generate or refine their own search terms before retrieving results, thereby leveraging the memory benefit of self-production even with minimal generative involvement.
Mirrors From index to page in the LlamaIndex primer.
SimilarityPostprocessor is a node postprocessor that applies a relevance floor to the top-k retrieved chunks, dropping any chunk with similarity score below 0.15. Without it, narrow-section queries would feed the LLM low-similarity noise, diluting faithfulness. The floor removes clear noise without starving the query.
Emotional arousal and humor works by giving arousing content consolidation priority through amygdala modulation of hippocampal encoding, with the retention advantage growing over a delay. A boundary condition is that the humorous element must be on the keyword-meaning link itself, not extraneous, and there is no evidence that sexual or gory content outperforms humorous or surprising content for delayed verbal-associative learning. In a retrieval pipeline, a Relevance scoring stage and similarity floor can apply this mechanism by scoring candidates for arousal while using the floor to keep humor tied to the link, optimizing for the delayed retention advantage.
Mirrors Relevance scoring and the similarity floor in the LlamaIndex primer.
The mechanism is implemented by the CodeSplitter class, which splits source code into chunks using the chunk_lines and chunk_lines_overlap parameters. The _code_splitter function returns a CodeSplitter instance that reads a source file and outputs overlapping chunks to preserve context between adjacent sections. Without the overlap, consecutive chunks would have no shared lines, breaking code references that span chunk boundaries.
The spacing effect's mechanism, grounded in Ebbinghaus and the Cepeda meta-analysis, shows that distributed practice consistently outperforms massed practice across ages and materials, with expanding schedules (e.g., 1, 3, 10, 30 days) now fit to individual forgetting curves. Boundary conditions highlight that scheduling dominates cue quality—spacing versus massing is a far larger effect than mnemonic improvements. Code-aware splitting applies this mechanism by breaking source code into logically spaced segments, ensuring review intervals are distributed rather than massed, thereby reducing forgetting through interleaved re-exposure.
Mirrors Code-aware splitting in the LlamaIndex primer.
RetrieverQueryEngine takes a retriever, an LLM, a response synthesizer built with the SYSDESIGN_QA template, and an optional reranker. It receives a query, retrieves relevant nodes, then uses the LLM to synthesize a spoken-clean answer from those nodes. Without it, the system would lack a retrieval-augmented generation pipeline for system-design questions, breaking the mechanism that grounds answers in the real codebase context.
Retrieval practice works by successful effortful retrieval strengthening and multiplying retrieval routes, unlike re-exposure. A boundary condition for mnemonics is that with practice, the learner should eventually retrieve meaning directly, letting the mnemonic scaffolding drop away. Asking the index, which retrieves stored information from memory on demand, directly applies this mechanism by forcing active recall rather than passive review.
Mirrors Asking the index in the LlamaIndex primer.
_code_grounding_violations takes a markdown string and a set of source identifiers. It extracts identifiers from fenced code blocks. If fewer than two thirds of those identifiers are whole tokens present in the source, it returns a list of violations. Without this check, hallucinated APIs or symbols could leak into the code excerpts, breaking the grounding guarantee that every snippet traces back to the real source.
The desirable difficulties principle holds that conditions slowing apparent learning enhance long-term retention because effortful processing builds storage strength distinct from retrieval strength, but difficulties are only desirable when the learner can overcome them; for beginners, excessive difficulty causes failed retrievals and frustration. In a retrieval pipeline, a 'Trust but verify' stage applies this mechanism by introducing a verification step that demands effortful retrieval or reasoning, thereby strengthening memory storage—provided the verification challenge remains surmountable for the learner's current level.
Mirrors Trust but verify in the LlamaIndex primer.
Chunking
The mechanism is implemented by the SentenceSplitter class and the CodeSplitter class, constructed via the helper function _code_splitter. These splitter objects take Document objects (each containing raw text and file metadata) and transform them into smaller LlamaIndex nodes. SentenceSplitter splits prose by sentence boundaries with configurable chunk size and overlap, while CodeSplitter uses a tree-sitter parser to split code files by logical function boundaries. Without these splitters, each source file or chapter narration would remain a single monolithic node, making fine-grained retrieval impossible and severely degrading the grounding quality of the generated explanations.
The chunking principle states working memory is limited to roughly four chunks, not bits, so recoding raw information into meaningful units multiplies capacity, as shown by Miller and Cowan. This mechanism applies beyond single vocabulary items, with acronyms and number-shape systems as examples. A retrieval pipeline that splits documents into chunks applies this same mechanism, breaking large texts into manageable units to fit memory constraints and enable efficient processing.
Mirrors Documents and chunks in the LlamaIndex primer.
The function build_index_from_nodes accepts nodes and an embed model, then calls sync_nodes to sync them to a stable Qdrant collection via a content‑keyed ref_doc_id, enabling incremental updates. It returns a scoped VectorStoreIndex. Without this mechanism, each run would re‑embed the entire corpus and orphan collections on every edit, causing redundant computation and index bloat.
The method of loci parasitizes spatial memory, which is evolutionarily old and high-capacity, with built-in ordering and cues from walking a route. Its boundary condition requires loci content to be user-anchored, such as their apartment or commute, or the built-in cues are lost. A nearest-neighbour vector space pipeline stage applies this mechanism by using spatial proximity in embedding space to retrieve ordered items, mimicking the route-based cues of loci.
Mirrors Nearest-neighbour vector space in the LlamaIndex primer.
The mechanism is implemented by build_index_from_nodes, which receives a list of TextNodes and an embed model (created by make_embed). It uses the embed model to generate vectors, syncs the nodes to a Qdrant collection while deduplicating unchanged content via a content-SHA manifest, and returns a scoped vector index. Without this, each grounding run would re-embed the entire corpus, losing incremental reuse.
The encoding specificity principle states that a cue only works at retrieval if it was part of the encoding, making cue immutability a hard invariant. Its mechanism, demonstrated by Tulving and Thomson (1973) and Godden and Baddeley (1975), shows that retrieval succeeds when the cue was encoded with the trace and fails otherwise, even for strong semantic associates. A boundary condition is that if a system regenerates mnemonics between sessions or A/B-tests cue variants, it severs the encoding-retrieval match, causing the mnemonic to silently stop working. A pipeline stage that performs Embeddings without an API puts this mechanism to work by using the same local embedding model and parameters at both encoding and retrieval, ensuring the query cue's embedding lands near the stored passage embeddings — mimicking retrieval-augmented generation's vector-space version of encoding specificity as described in the source.
Mirrors Embeddings without an API in the LlamaIndex primer.
In a retrieval-augmented generation pipeline, encoding quality (principles 1–7) does the heavy lifting by determining the initial strength of each memory trace, but without retrieval practice (9–10) and spacing (8) that strength fails to compound over delays, producing only short-term gains. Specificity (13) then acts as the invariant that keeps the whole retrieval loop connected. A pipeline that implements these together outperforms a piecemeal approach because nailing 1–7 alone yields impressive demos but poor retention at month three—the actual lesson from the keyword-method literature is that encoding techniques need a spaced-retrieval schedule. The interaction constraint on generation is a high-value single rule, and spaced-retrieval integration determines whether the product truly works long-term.
References
- Atkinson, R. C., & Raugh, M. R. (1975). An application of the mnemonic keyword method to the acquisition of a Russian vocabulary. Journal of Experimental Psychology: Human Learning and Memory, 104(2). DOI: 10.1037/0278-7393.1.2.126
The original keyword-method studies — the paradigm every keyword mnemonic descends from.
- Atzert, C., Děchtěrenko, F., Lukavský, J., & Busch, N. A. (2026). Scene memorability reflects representational distinctiveness within visual categories. bioRxiv. DOI: 10.64898/2026.03.20.713124
Scene memorability tracks how distinct an item's representation is within its own category — distinctiveness made measurable. Preprint.
- Balepur, N., Shu, M., Hoyle, A., Robey, A., Feng, S., Goldfarb-Tarrant, S., & Boyd-Graber, J. (2024). A SMART Mnemonic Sounds like “Glue Tonic”: Mixing LLMs with Student Feedback to Make Mnemonic Learning Stick. arXiv:2406.15352.
Aligns an LLM mnemonic generator with student feedback on which mnemonics actually help learning. Preprint.
- Bengio, Y., Louradour, J., Collobert, R., & Weston, J. (2009). Curriculum learning. ICML 2009. DOI: 10.1145/1553374.1553380
Orders training examples easy-to-hard — difficulty as a schedulable training variable.
- Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings. In J. Metcalfe & A. Shimamura (Eds.), Metacognition: Knowing about knowing. MIT Press.
The desirable-difficulties framework: retrieval strength and storage strength are distinct.
- Brysbaert, M., Warriner, A. B., & Kuperman, V. (2014). Concreteness ratings for 40 thousand generally known English word lemmas. Behavior Research Methods, 46. DOI: 10.3758/s13428-013-0403-5
Concreteness norms covering ~40k English words — imageability is measurable at scale.
- Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T., & Rohrer, D. (2006). Distributed practice in verbal recall tasks: A review and quantitative synthesis. Psychological Bulletin, 132(3). DOI: 10.1037/0033-2909.132.3.354
839 assessments: distributed practice beats massed practice everywhere, at every age, for every material type tested.
- Cepeda, N. J., Vul, E., Rohrer, D., Wixted, J. T., & Pashler, H. (2008). Spacing effects in learning: A temporal ridgeline of optimal retention. Psychological Science, 19(11). DOI: 10.1111/j.1467-9280.2008.02209.x
The optimal gap scales with the retention interval — roughly 10–20% of the time until you need the memory.
- Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24(1). DOI: 10.1017/S0140525X01003922
The modern revision: ~4 chunks, not 7.
- Craik, F. I. M., & Lockhart, R. S. (1972). Levels of processing: A framework for memory research. Journal of Verbal Learning and Verbal Behavior, 11(6). DOI: 10.1016/S0022-5371(72)80001-X
Memory as a byproduct of processing depth: semantic beats phonemic beats orthographic.
- Craik, F. I. M., & Tulving, E. (1975). Depth of processing and the retention of words in episodic memory. Journal of Experimental Psychology: General, 104(3). DOI: 10.1037/0096-3445.104.3.268
The congruity refinement: elaboration helps most when it forms an integrated structure around the target.
- Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. NAACL 2019 (arXiv:1810.04805). DOI: 10.18653/v1/N19-1423
Masked-language-model pretraining is unsatisfiable with surface features — it forces deep semantic encoding.
- Dresler, M., Shirer, W. R., Konrad, B. N., et al. (2017). Mnemonic training reshapes brain networks to support superior memory. Neuron, 93(5). DOI: 10.1016/j.neuron.2017.02.003
Six weeks of loci training in novices durably improved recall and shifted brain connectivity toward the athlete pattern.
- Ebbinghaus, H. (1885). Über das Gedächtnis: Untersuchungen zur experimentellen Psychologie. Duncker & Humblot, Leipzig.
The forgetting curve and the first demonstration of distributed practice.
- Eldho Paul, A., & Sunar, M. S. (2026). A Dual-Conditioned CVAE for Interpretable Automatic Question Generation: Controlling Topic and Difficulty through Latent Spaces. Proceedings of the 18th International Conference on Computer Supported Education (CSEDU). DOI: 10.5220/0015091000004021
Generates assessment questions with independent topic and difficulty control — the item-supply engine for retrieval practice, with a difficulty knob for the effort gradient.
- Elsner, M., Chen, J., & Sims, A. D. (2026). Memory retrieval as a pressure towards chunking in morphological inflection: the case of Turkish. Linguistics Vanguard. DOI: 10.1515/lingvan-2024-0235
An RL-optimized model where memory-retrieval pressure drives recoding of morphological sequences into multi-affix chunks — chunking emerging as an optimization, not a hand-coded rule. Computational model.
- Ericsson, K. A., Chase, W. G., & Faloon, S. (1980). Acquisition of a memory skill. Science, 208(4448). DOI: 10.1126/science.7375930
SF extended digit span from 7 to ~80 by chunking digits into running times.
- Godden, D. R., & Baddeley, A. D. (1975). Context-dependent memory in two natural environments: On land and underwater. British Journal of Psychology, 66(3). DOI: 10.1111/j.2044-8295.1975.tb01468.x
Divers recalled underwater-learned words better underwater — the famous context-effect demonstration.
- Graves, A., Wayne, G., Reynolds, M., Harley, T., et al. (2016). Hybrid computing using a neural network with dynamic external memory. Nature, 538. DOI: 10.1038/nature20101
The DNC stores items at addressable memory locations and a temporal-link matrix records write order — traversable like a route.
- Hunt, R. R., & McDaniel, M. A. (1993). The enigma of organization and distinctiveness. Journal of Memory and Language, 32(4). DOI: 10.1006/jmla.1993.1023
Item-specific vs. relational processing — best recall comes from doing both.
- Hwang, S., Seo, J., Kim, S., & Lee, J. (2026). A Multi-Agent Framework for Feature-Constrained Difficulty Control in Reading Comprehension Item Generation. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Long Papers). DOI: 10.18653/v1/2026.acl-long.1267
Generates items at a targeted difficulty via feature constraints — the knob that makes difficulty a tunable, desirable one.
- Kang, S., Gwon, M., Kwon, S. Y., Lee, J., Lan, A., Raj, B., & Singh, R. (2025). PhoniTale: Phonologically Grounded Mnemonic Generation for Typologically Distant Language Pairs. EMNLP 2025 (arXiv:2507.05444).
Retrieves an L1 keyword sequence by IPA-level phonological alignment to the L2 word, so the recall cue is grounded in the sound overlap present at encoding.
- Karpicke, J. D., & Blunt, J. R. (2011). Retrieval practice produces more learning than elaborative studying with concept mapping. Science, 331(6018). DOI: 10.1126/science.1199327
Retrieval practice beats even elaborative concept mapping.
- Kleinsmith, L. J., & Kaplan, S. (1963). Paired-associate learning as a function of arousal and interpolated interval. Journal of Experimental Psychology, 65(2). DOI: 10.1037/h0045203
Arousing pairs were recalled worse than neutral at immediate test but much better a week later.
- Koneru, S., & Powell, B. (2026). Visual Linguistics of Memorability: A Caption-Based Framework for Image Memorability Analysis. Proceedings of the 2026 ACM Southeast Conference. DOI: 10.1145/3746467.3801518
Ties image memorability to linguistic features of captions — memorability as a text-measurable property.
- Lee, J., & Lan, A. (2023). SmartPhone: Exploring Keyword Mnemonic with Auto-generated Verbal and Visual Cues. arXiv:2305.10436 (AIED 2023).
End-to-end keyword-mnemonic pipeline: an LLM picks the keyword and writes the verbal cue, a text-to-image model renders the scene. Preprint.
- Lee, J., McNichols, H., & Lan, A. (2024). Exploring Automated Keyword Mnemonics Generation with Large Language Models via Overgenerate-and-Rank. arXiv:2409.13952.
Overgenerate many candidate keyword mnemonics, then rank on psycholinguistic measures — the scoring stage a diversity penalty plugs into. Preprint.
- Lee, J., Park, S., Lee, K., Hong, J., Yoshimura, K., & Lan, A. (2026). Personalizing Kanji Memorization: Designing Adaptive Mnemonics Based on Learner Preferences Using Large Language Models. Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems. DOI: 10.1145/3772363.3798551
Adapts generated mnemonics to each learner's preferences — the personalization surface that lets support scale down as items mature.
- Lee, J., Scarlatos, A., & Lan, A. (2025). Interpretable Mnemonic Generation for Kanji Learning via Expectation-Maximization. EMNLP 2025 (arXiv:2507.05137).
Decomposes each kanji into its components and generates a mnemonic per component via an EM algorithm — chunking as an explicit generation step.
- Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020 (arXiv:2005.11401).
Encodes a query into a cue that must land near stored passage embeddings for retrieval to succeed.
- Maguire, E. A., Valentine, E. R., Wilding, J. M., & Kapur, N. (2002). Routes to remembering: The brains behind superior memory. Nature Neuroscience, 6(1). DOI: 10.1038/nn988
Memory champions have ordinary capacity but reorganized, spatially-anchored retrieval strategies.
- McDaniel, M. A., & Einstein, G. O. (1986). Bizarre imagery as an effective memory aid: The importance of distinctiveness. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12(1). DOI: 10.1037/0278-7393.12.1.54
The bizarreness advantage appears in mixed lists and evaporates when every item is bizarre.
- Miller, G. A. (1956). The magical number seven, plus or minus two: Some limits on our capacity for processing information. Psychological Review, 63(2). DOI: 10.1037/h0043158
Working memory is limited in chunks, not bits.
- Paivio, A. (1991). Dual coding theory: Retrospect and current status. Canadian Journal of Psychology, 45(3). DOI: 10.1037/h0084295
The canonical statement of dual coding: concrete content is encoded in independent verbal and imaginal systems.
- Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., et al. (2021). Learning Transferable Visual Models From Natural Language Supervision. ICML 2021 (arXiv:2103.00020).
CLIP aligns an image channel and a text channel in one shared space — two encodings of the same content.
- Ramsauer, H., Schäfl, B., Lehner, J., Seidl, P., et al. (2020). Hopfield Networks is All You Need. ICLR 2021 (arXiv:2008.02217).
A modern Hopfield layer completes a partial cue to a stored pattern in one update — retrieval as pattern completion.
- Robinson, J., Chuang, C.-Y., Sra, S., & Jegelka, S. (2020). Contrastive Learning with Hard Negative Samples. ICLR 2021 (arXiv:2010.04592).
Up-weights the most confusable negatives and pushes their representations apart.
- Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3). DOI: 10.1111/j.1467-9280.2006.01693.x
After a week, retrieval practice retained ~50% more than restudy — even though restudy felt more effective.
- Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T. P., & Wayne, G. (2019). Experience Replay for Continual Learning. NeurIPS 2019 (arXiv:1811.11682).
Re-exposes a network to buffered past experiences interleaved with new learning to counter catastrophic forgetting.
- Roy, A., Wang, Q., & MacLellan, C. (2026). Grounded Concreteness: Human-Like Concreteness Sensitivity in Vision–Language Models. Findings of the Association for Computational Linguistics: ACL 2026. DOI: 10.18653/v1/2026.findings-acl.2081
Vision–language models track human concreteness ratings, making automatic concreteness scoring practical beyond word norms.
- Santoro, A., Raposo, D., Barrett, D. G. T., Malinowski, M., et al. (2017). A simple neural network module for relational reasoning. NeurIPS 2017 (arXiv:1706.01427).
A Relation Network fuses each pair of entities through a shared MLP into one joint relational representation.
- Schaul, T., Quan, J., Antonoglou, I., & Silver, D. (2015). Prioritized Experience Replay. ICLR 2016 (arXiv:1511.05952).
Replays transitions with probability proportional to TD-error — the most surprising experiences are replayed most.
- Schmidt, S. R. (1994). Effects of humor on sentence memory. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20(4). DOI: 10.1037/0278-7393.20.4.953
Humorous sentences are recalled better than matched non-humorous ones — with the same mixed-list caveat as bizarreness.
- Sennrich, R., Haddow, B., & Birch, A. (2016). Neural Machine Translation of Rare Words with Subword Units. ACL 2016 (arXiv:1508.07909).
Byte-pair encoding merges frequent adjacent symbols into reusable subword units — recoding into chunks.
- Settles, B., & Meeder, B. (2016). A trainable spaced repetition model for language learning. Proceedings of ACL 2016. DOI: 10.18653/v1/P16-1174
Duolingo's half-life regression — fitting individual forgetting curves per item per user.
- Shao, Y., Xiong, Z., Wu, Z., Wang, S., Zhou, Y., Ouyang, Y., Tao, Y., & Li, Q. (2026). WordCraft: Scaffolding the Keyword Method for L2 Vocabulary Learning with Multimodal LLMs. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. DOI: 10.1145/3772318.3790668
Multimodal LLM scaffolding for learner-created keyword mnemonics — the system assists, the learner generates.
- Shu, R., Wang, X., & Hardy, I. (2026). Measuring Optimal Challenge: Trajectory-Based Difficulty Alignment in Open-Ended Language Tutoring. Proceedings of the 21st Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2026). DOI: 10.18653/v1/2026.bea-1.46
Aligns tutoring difficulty to the learner's trajectory, targeting the optimal-challenge zone — difficulty held at the edge the learner can just overcome.
- Slamecka, N. J., & Graf, P. (1978). The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory, 4(6). DOI: 10.1037/0278-7393.4.6.592
Material you produce yourself — even trivially — is remembered better than material you receive.
- Spens, E., & Burgess, N. (2024). Hippocampo-neocortical interaction as compressive retrieval-augmented generation. bioRxiv. DOI: 10.1101/2024.11.04.621950
A computational model where compressed memories are replayed to train a generative network — spacing and consolidation implemented in silico. Preprint; model-level analogy.
- Tulving, E., & Thomson, D. M. (1973). Encoding specificity and retrieval processes in episodic memory. Psychological Review, 80(5). DOI: 10.1037/h0020071
A cue works at retrieval only to the extent it was part of the encoding.
- von Restorff, H. (1933). Über die Wirkung von Bereichsbildungen im Spurenfeld. Psychologische Forschung, 18. DOI: 10.1007/BF02409636
The original isolation effect: an item that differs from its surroundings is recalled dramatically better.
- Wollen, K. A., Weber, A., & Lowry, D. H. (1972). Bizarreness versus interaction of mental images as determinants of learning. Cognitive Psychology, 3(3). DOI: 10.1016/0010-0285(72)90020-5
Interacting images beat separated images — and interaction mattered more than bizarreness.
- Wu, W., & Smith, D. A. (2023). Composition and Deformance: Measuring Imageability with a Text-to-Image Model. Proceedings of the 5th Workshop on Narrative Understanding, ACL. DOI: 10.18653/v1/2023.wnu-1.16
Scores sentence-level imageability by generating images from the text and measuring generation quality and consistency.
- Wulff, C., Kruse, L., & Steinicke, F. (2026). Enhancing Memory Recall Through AI-Assisted Method of Loci in Virtual Reality. Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems. DOI: 10.1145/3772363.3798815
Builds and evaluates a generated memory palace in VR — an object placed at each locus, a scene generated per pair, tested at immediate and one-week recall.
- Xiao, Q., & Wang, J. (2024). DRL-SRS: A Deep Reinforcement Learning Approach for Optimizing Spaced Repetition Scheduling. Applied Sciences, 14(13). DOI: 10.3390/app14135591
A Transformer + deep-RL scheduler that estimates recall probability and optimizes review timing — the learned-scheduler lineage pushed past regression fits. Computational model.
- Zelikman, E., Wu, Y., Mu, J., & Goodman, N. D. (2022). STaR: Bootstrapping Reasoning With Reasoning. NeurIPS 2022 (arXiv:2203.14465).
Fine-tunes a model on the correct chain-of-thought rationales it generated itself.