The existing memory systems, in detail

A reference for the systems Akasha borrows from, what each was built to do, and where each stops being useful. Companion to memory-system.md, which describes what this repo actually implements.

The division of labour

The single most useful thing to know is that these systems are not competing. They solve different problems and stack cleanly:

A competitive memoriser uses four of these at once and calls the combination "my system". Most people arguing about which technique is best are comparing things from different rows.

Akasha's allocation: loci for structure, Major for numbers, link/story for individual facts, contrast for confusables, and its own scheduler on top.

1. Encoding: link and story

The oldest and least formal. Convert each item into a concrete image, then associate images in a chain — item one collides with item two, which collides with item three.

Why it works. Three effects compound. Dual coding (Paivio): a fact stored as both words and imagery has two retrieval routes. Elaborative encoding: the effort of building the association creates the connective tissue that retrieval later travels along. Bizarreness: unusual images are better recalled than mundane ones, though the effect is smaller than the popular literature claims and mostly disappears when everything is bizarre.

Where it fails. A chain is fragile — lose one link and everything after it goes with it. There is no random access: to reach item nine you walk one through eight. And it gives you no way to know you have finished.

In Akasha. Every peg field. It is the workhorse, and its weaknesses are exactly what loci fix.

2. Structure: the method of loci

The central technique of the whole tradition.

The founding story. Simonides of Ceos, around 500 BC, leaves a banquet moments before the roof collapses. The bodies are unidentifiable, but he can name every victim by recalling where each had been sitting. The lesson drawn: spatial position is a retrieval index. Cicero tells the story in De Oratore; the fullest surviving classical instruction is in the anonymous Rhetorica ad Herennium (c. 86–82 BC), which distinguishes memory for things from memory for words and is blunt that the second is much harder.

How it works. Choose a place you know well. Fix an ordered route through it — a sequence of distinct positions (loci). Place one image at each. To recall, walk the route.

Why it works. Spatial memory is unusually robust and effortless: you can walk your childhood home mentally without having tried to memorise it. Loci borrow that scaffolding for material that has none. And the route imposes order and completeness on an unordered set — you can tell you have missed one, which no other technique gives you.

Variants worth knowing.

• Roman room — one room, objects within it as loci. Small, quick to build.
• Journey method — Dominic O'Brien's name for a route through real places rather than a building. Scales better; a walk to the shops has dozens of loci and needs no invention.
• Memory theatre — Giulio Camillo's 16th-century project: an architectural structure where position encodes meaning, not just order, so the building is a map of knowledge rather than a filing cabinet.
• Ars memorativa — Ramon Llull's combinatorial wheels (13th c.) and Giordano Bruno's systems (16th c.). Attempts to make memory generative rather than merely retentive. Intellectually extraordinary, practically unusable.
• Matteo Ricci — the Jesuit who taught loci in Ming China in the 1590s as a way in to the examination elite. Jonathan Spence's The Memory Palace of Matteo Ricci is the book.

The modern evidence. Dresler et al. (2017) trained non-experts in loci for six weeks and got large, durable gains along with brain connectivity patterns shifting toward those of memory athletes. It is one of the better-supported findings in the area: this is a trainable skill, not a talent.

Where it fails. Palaces cost real effort to build and maintain, and reusing one for new material can interfere with the old. It suits ordered material and is overkill for isolated pairs.

In Akasha. Two places. The era clock — twelve eras, twelve hours each, 144 addressed slots. And the country route, where the map itself is the palace: Europe is already an ordered spatial structure you half-know, so the loci come free.

3. Numbers: peg systems

Numbers are the hardest material because digits have no imagery. Peg systems pre-assign one fixed image per number, learned once and reused forever.

• Number-shape — the image resembles the glyph. 1 candle, 2 swan, 3 handcuffs, 4 sailboat, 7 cliff, 8 snowman. Instant to learn, caps out around 10.
• Number-rhyme — 1 bun, 2 shoe, 3 tree, 4 door, 5 hive. Same range, same ease. Good for a first ordered list of ten.
• Alphabet peg — A ape, B bee, C sea. Twenty-six ordered slots, useful where the ordering is alphabetical anyway.

Where they fail. They do not scale. Ten or twenty-six fixed pegs handles a shopping list, not a phone number, and reusing the same pegs for two lists at once causes interference. The phonetic systems below exist to break that ceiling.

4. Numbers: the Major System

The most important number system, and the one Akasha uses.

The mapping — digits to consonant sounds, vowels free:

Two rules people get wrong. It is sounds, not letters — "cough" is 78 (k-f), not 707. And vowels plus w, h, y are free filler, which is what gives you room to build real words.

So 34 can be MOWER, MARE, MIRROR or EMIR; all are m-r. 1789 splits as 17 and 89 — TACK and VIP.

History. Pierre Hérigone gives the earliest known digit-to-consonant scheme in his Cursus mathematicus (1634). Stanislaus Mink von Wennsshein — a pseudonym — refines it in 1648. Richard Grey's Memoria Technica (1730) is a competing English scheme using vowels as well, and Lewis Carroll later built his own variant to memorise logarithms. Aimé Paris systematises the modern phonetic form in the early 1800s. The name is generally attributed to Major Beniowski, whose Anti-Absurd or Phrenotypic Alphabet appeared in 1842, though the attribution is loose. Harry Lorayne brought it to a mass audience in the 20th century.

The ancient cousin. Katapayadi, from India, maps consonant groups to digits and was used to embed astronomical constants and the names of Carnatic melakarta ragas into verse — so a raga's name tells a trained musician its scale. Independently invented, centuries earlier, and arguably more elegant because the encoded numbers hide inside meaningful words. Abjad numerals and Hebrew gematria are related in mechanism but were mostly used in the other direction: extracting numbers from words rather than storing numbers as them.

The one artifact worth building — and it is now built. A fixed 00–99 peg list, one settled image per two-digit pair, plus the single digits: 110 images in strands/system/. After that every number is chunked into pairs and read off, with no on-the-fly word-hunting. This is the highest-leverage thing in the whole tradition.

Every code is machine-verified: tests/test_major.py re-derives all 110 from their spellings and fails if a word does not encode to the number it is filed under. Where a peg needs a pronunciation exception — silent letters in GNOME, KNIFE, COMB, LAMB, TOMB; the affricate in MATCH and NOTCH — that exception is listed explicitly rather than fudged, because a peg whose reading is ambiguous is a peg that will fail you under pressure.

Where it fails. It is only for numbers, and it needs the peg list built before it pays. Improvising words mid-recall is slow and error-prone, which is exactly why the fixed list matters.

In Akasha. Every population and area peg, and every date in the era strand. Note the deliberate design: because the era supplies the century, date pegs only carry two digits, so the entire history spine needs only the 00–99 list and nothing longer.

5. Numbers: Dominic, PAO, and beyond

Extensions for people who need more than 110 images.

• Dominic System (Dominic O'Brien, eight-time World Memory Champion). Digits map to letters (1 A, 2 B, 3 C, 4 D, 5 E, 6 S, 7 G, 8 H, 9 N, 0 O), and each two-digit pair becomes a person with a characteristic action. 43 is DC, so David Copperfield, vanishing. Many find people far more memorable than objects, which is the whole argument for it over Major.
• PAO (Person–Action–Object) — every number 00–99 gets a fixed person, a fixed action and a fixed object. Three pairs then compress into one scene: person of the first doing the action of the second to the object of the third. Six digits per image. Standard equipment at competition level.
• Ben System (Ben Pridmore) — one image per number 000–999. A thousand images memorised cold, for speed cards and long-digit events.
• Shadow System (Johannes Mallow) — another three-digit system at the same extreme.

Where they fail for our purposes. All of them are built for long strings of digits under time pressure. Akasha has populations, areas and years — two to four digits, no clock. Major alone is the right tool, and PAO would be several hundred hours of setup for material that never needs it.

The one thing worth stealing is the person-per-item idea. Assigning each country a person is what would let multi-country events — wars, treaties, partitions — be encoded as scenes with recognisable actors. That is the missing bridge between the countries strand and the history strand.

6. Retrieval structures and chunking

Less famous than palaces, and arguably more important.

Miller's "magical number seven" (1956) set working memory at roughly seven chunks — but a chunk is whatever you have made into one unit. Chase and Ericsson's work in the early 1980s followed a runner, "SF", who reached 80+ digits not by expanding memory but by recoding digit strings as running times he already knew, then organising those into a hierarchy. They called the resulting scaffold a retrieval structure.

That is the actual principle behind everything above: expertise is not more memory, it is better addresses.

Two-stage retrieval is the practical form. Rather than searching 144 events, you retrieve the era (one of twelve), then the hour (one of twelve). Two easy lookups beat one hard one, and the coarse stage is more robust — you will remember that something is late-Revolutions long after you have lost the year.

In Akasha. The era spine is a retrieval structure. So is the country route. This is why both are worth more than the pegs hanging off them.

7. Discrimination

Not a classical named system, but it addresses the failure mode that actually occurs.

You will not forget Slovakia's flag. You will confuse it with Slovenia's. Similar items interfere; the more you learn, the worse it gets, because each new near-neighbour is another competitor at retrieval time. Studying items in isolation actively encourages this, because you never practise telling them apart.

The fix is to train the discriminating feature directly: interleave the confusable set, and make the question "which of these is it, and how do you know" rather than "what is this". This is a special case of Robert Bjork's desirable difficulties — interleaving and contrast feel worse during study and produce better retention.

In Akasha. The contrast atoms, and the entire arts strand.

8. Scheduling

Orthogonal to everything above. These decide when, not how.

• Ebbinghaus (1885) — the original forgetting curve, measured on himself with nonsense syllables. Retention decays roughly logarithmically, and each successful review flattens the curve.
• The testing effect — retrieval practice beats re-reading, substantially (Roediger and Karpicke, 2006). Being tested is not measurement; it is the learning event. This is why Akasha has no "read through" mode.
• The generation effect — material you produce yourself is retained better than material you are given (Slamecka and Graf, 1978). The reason self-written pegs beat supplied ones, and the reason akasha peg exists.
• The spacing effect — the same total study time produces more retention when distributed. Optimal gap scales with how long you want to retain.
• Leitner box (1970s) — physical cards in numbered boxes; correct moves a card back a box, wrong sends it to the front. Spaced repetition with no arithmetic.
• SM-2 (Woźniak, 1987) — SuperMemo's published algorithm and the basis of Anki's classic scheduler. Each card has an ease factor adjusted by grade; intervals multiply. Simple, and it has run the field for thirty years.
• FSRS — models memory with three variables: difficulty, stability (how long until recall probability falls to the target) and retrievability (probability right now), with parameters fitted to large review datasets. Better calibrated than SM-2, particularly on lapses.

In Akasha. akasha/scheduler.py uses FSRS's shape — stability, difficulty, and the power-law curve R(t) = (1 + t/9S)^-1 — with hand-chosen constants rather than fitted weights, stated plainly rather than passed off. Swapping in real py-fsrs is one function.

9. What to use when

10. Applied to Akasha

Two things to build, in this order.

1. A 00–99 Major peg list. Done — strands/system/, 110 pegs, all verified. Drill it first: akasha review --strand system. It is the prerequisite for the population, area and date material being fast rather than laborious, and it is reusable for the rest of your life.
2. The Europe route. A fixed walk over the 46 countries. Costs almost nothing because the material already is the structure, and it is the only thing that will tell you when you have missed one.

Everything else — PAO, three-digit systems, Llull's wheels — is either solving a problem Akasha does not have, or is best enjoyed as history.

Reading

• **Frances Yates, The Art of Memory (1966)** — the standard scholarly history, from Simonides through Camillo, Llull and Bruno. Dense, superb.
• ***Rhetorica ad Herennium*, Book III** — the classical source, and still clearer instruction than most modern writing on loci.
• **Joshua Foer, Moonwalking with Einstein (2011)** — the popular account that revived interest; a journalist trains for the US Memory Championship.
• **Jonathan Spence, The Memory Palace of Matteo Ricci (1984)**.
• Dominic O'Brien, on the Dominic System from its inventor.
• artofmemory.com — the best practical modern resource, including ready-made 00–99 Major lists if you would rather adapt than invent.