Genealogy
Fifty thousand years of retention.
Memory was not invented in 1949 with Hebb, or in 1972 with Tulving, or in 2014 with the Neural Turing Machine. It was solved across many substrates, by many cultures, over fifty millennia. The table classifies what was already there.
Aboriginal songlines
Continental-scale knowledge transmission across at least fifty millennia, encoded into song, dance, petroglyph, ceremony, and place. The deepest verified case of multimodal cross-validated retention. The substrate is the landscape itself.
Inca khipu (Caral predecessors)
Knotted-cord record-keeping in the Andes, decimal-positional and categorical, hierarchical (primary → pendant → subsidiary). Specialist class (quipucamayoc) trained in yachaywasi. Cross-encoder verification by independent reading.
Vedic recitation (Saṃhitā era)
Pada-, krama-, jaṭā-, ghana-pāṭha: nested permutation recitation. Syllable-exact preservation of tens of thousands of verses across three millennia, verified against a 19th-century global variorum showing ~12 variants. Random-access via avadhāna.
The method of loci
The technique that became De Oratore and Rhetorica ad Herennium. Visual imagery bound to imagined spatial loci. Modern fMRI work confirms it exploits the hippocampal place-cell substrate that O'Keefe and Nadel formalized in 1978.
al-quwwa al-hāfiẓa
The retentive faculty as one of five internal senses, distinct in function and locus from estimation and imagination. The first systematic Western-influencing taxonomy of memory subsystems. Inherited into Aquinas and the medieval art of memory.
Über das Gedächtnis
The first quantitative measurement of forgetting. Method of savings on nonsense syllables. The forgetting curve, the first measured retention time-constant in psychology. Replicated by Murre and Dros in 2015 with a refined multi-exponential fit.
Remembering
Recall as reconstruction, not retrieval. The schema-driven nature of remembering shown by serial-reproduction experiments with "The War of the Ghosts." The empirical basis for what L7 of the cascade is.
EDVAC draft
The stored-program memory architecture. The first explicit unification of program and data into a single addressable substrate. Made the entire L2-tier indexed-lookup family the default for computing.
The Organization of Behavior
Neurons that fire together wire together. The proposed mechanism for biological write at the synaptic substrate. The W₂ axis value, twenty-four years before its molecular substrate (LTP) was found.
Magnetic core memory
The first non-volatile random-access memory at scale. Destructive read, mandatory rewrite. The first widely deployed substrate where recall was already a write event.
The magical number seven
The first formal capacity claim for working memory. Refined by Cowan in 2001 to four-plus-or-minus-one for chunks not supplemented by long-term memory.
The case of H.M.
Bilateral hippocampectomy → anterograde amnesia with preserved procedural learning. The first hard demonstration that declarative and non-declarative memory live on different substrates. The empirical foundation for Squire's later taxonomy.
The thermodynamic floor
Erasing a bit dissipates at least k_B T · ln 2 joules. The Eᵂ axis bottom, eleven years before its first experimental verification. The "+0" against which every other substrate is plotted.
DRAM and refresh-as-architecture
One transistor, one capacitor. Leakage made an axis: the ρ ≫ 1 substrate that defined what an "active refresh process" is in silicon. Every modern dynamic memory is the same idea.
The modal model
Sensory register → short-term store → long-term store, with control processes shuttling between. The first systematic decomposition of biological memory into a tiered cascade. Refined by Baddeley and Hitch in 1974.
Episodic vs. semantic
The taxonomy of long-term memory. Episodic, semantic, procedural — each with distinct phenomenology and (it was later shown) distinct substrates. The R axis got its values from this work.
LTP
Long-term potentiation in the rabbit hippocampus. The molecular substrate for the Hebbian write proposed twenty-four years earlier. Later resolved into an early phase (hours, no protein synthesis) and a late phase (weeks to lifetime, protein synthesis).
Encoding specificity
Recall succeeds to the extent the retrieval context overlaps the encoding context. The empirical basis for the A₅ context-reinstatement axis value. Made cognitive memory irreducibly contextual.
The Hippocampus as a Cognitive Map
Place cells. The spatial substrate of episodic memory. The biological reason the method of loci works, identified twenty-five centuries after Simonides.
Attractor networks
Content-addressable memory as phase-space flow over an energy landscape. Capacity bounded at 0.138 N (Amit, Gutfreund, Sompolinsky, 1985). The cleanest example of a substrate whose C axis is a hard phase transition.
Sparse distributed memory
High-dimensional content-addressable memory with capacity independent of vector dimension. The mathematical kin of what would later be called hyperdimensional computing. Rediscovered as the right model for transformer attention by Bricken and Pehlevan in 2021.
CRISPR sequences (function 2007)
Repeated DNA arrays in E. coli, function unknown. In 2007, Barrangou et al. showed the spacers between repeats are an adaptive immune memory — bacteria recording past viral encounters as DNA. The deepest write-acquisition substrate on Earth.
Complementary Learning Systems
Hippocampus = fast, sparse, episodic; neocortex = slow, distributed, semantic. Sleep-driven consolidation moves traces between them. The two-substrate architecture that prevents catastrophic forgetting in biology.
LSTM
Gated memory cells in recurrent networks. Solved the vanishing-gradient problem by making the write regime explicitly gated. The first neural network with a coherent τ axis.
Reconsolidation
Recall destabilizes the trace. Protein-synthesis inhibitors injected during recall — not during encoding — erase the memory. Established that recall is a write event in biological memory. The F₄ axis value.
Landauer verified
Experimental verification of Landauer's principle in a colloidal-particle double-well bit. The first measurement of the +0 floor. Hong et al. confirmed it in nanomagnetic bits four years later.
DNA archival storage
Demonstrated 5.5 petabits per cubic millimeter (Church) and 2.2 petabytes per gram (Goldman). The molecular substrate at the millennium-τ corner of the table.
Neural Turing Machine
Differentiable external memory with attentional addressing. The first deep-learning attempt at a memory-augmented architecture that treats retention as a first-class axis. Extended in 2016 to the Differentiable Neural Computer.
4.7 bits per synapse
Nanoconnectomic measurement of synaptic spine sizes shows 26 distinct strength states — 4.7 bits of information storage per synapse. An order of magnitude above prior estimates. The synaptic C axis got its first hard number.
The transformer KV cache
Attention is all you need. The KV cache is a working-memory tier that the field built and never named, sized, or budgeted. The substrate that put L7-generative-reconstruction at the top of every modern query.
RAG
Retrieval-augmented generation. The first widely adopted external semantic store for LLMs. Acknowledged the field had a memory problem; chose to externalize rather than build the cascade properly.
The Periodic Stack of Memory
Fifty thousand years of retention. The map.
The reading
Fifty thousand years, one map.
Each era added a substrate, a mechanism, or a measurement that fills another cell on the table. The genealogy is the history of cells finding their coordinates.