Eleven orthogonal dimensions.

01 · HOW LONG IT LIVES

half-life of the trace, in seconds

Retention

The decay half-life of the trace, measured. Continuous on a log scale that spans 10⁻⁹ s at the register level to 10¹⁷ s for DNA across geological time. Working memory falls without rehearsal in fifteen to thirty seconds. DRAM cells leak in sixty-four milliseconds. Early-phase LTP holds for hours; late-phase, weeks to a lifetime. NAND flash specs ten years. Vedic recitation has held three thousand. Aboriginal songlines, fifty thousand. Discretized to eight orders-of-magnitude bands.

Admissible classes

Citations

Ebbinghaus, H. Über das Gedächtnis (1885). · Murre & Dros, PLOS ONE 10 (2015) — multi-exponential fit. · Peterson & Peterson, JEP 58 (1959) — working-memory decay.

02 · INFO LIFE OVER SUBSTRATE LIFE

τ_information / τ_substrate · dimensionless

Refresh ratio

The ratio of how long the information persists to how long the substrate carrying it persists. When the ratio is one, the trace and its substrate die together — write-once persistence. When the ratio is much greater than one, the information must be actively refreshed onto fresh substrate. The same pattern recurs across radically different substrates: DRAM charge decays in milliseconds, information persists indefinitely via electronic refresh. Memory B-cell populations sustain antibody memory across decades while individual cells turn over in months. Vedic recitation has carried the Saṃhitā across three millennia on a substrate (the human teacher) that lives seventy years. The ratio names the recurring trick. Lose the refresh process and the information is gone in one τ_substrate.

Admissible classes

Citations

Dennard, US Patent 3,387,286 (1968) — refresh-as-architecture. · von Neumann, Probabilistic Logics… (1956). · Hammerlund et al., Nat Med 9 (2003). · Staal, Nature 322 (1986) — Vedic transmission fidelity.

03 · STATES BEFORE INTERFERENCE

distinguishable retained states · bits

Capacity

The number of distinguishable retained states before interference destroys retrieval — not raw size, retrievable size. Hopfield's classical network holds 0.138 N patterns before a first-order phase transition erases all of them at once. Working memory holds four-plus-or-minus-one chunks. A single hippocampal synapse encodes 4.7 bits across 26 measurable strength states. DNA stores 2.2 petabytes per gram. The axis splits between substrates whose capacity fails gracefully and those whose capacity is a hard wall.

Admissible classes

Citations

Hopfield, PNAS 79 (1982). · Amit, Gutfreund & Sompolinsky, PRL 55 (1985). · McEliece et al., IEEE TIT 33 (1987). · Miller, Psychol. Rev. 63 (1956). · Cowan, BBS 24 (2001). · Bartol et al., eLife 4 (2015). · Goldman et al., Nature 494 (2013).

04 · COST TO WRITE

joules per bit · log₁₀(E / Landauer) above floor

Write energy

The energy dissipated to store one bit, measured per write. Landauer's principle sets the floor: k_B T · ln 2 ≈ 2.87 × 10⁻²¹ J/bit at 300 K, verified experimentally by Bérut (2012) and Hong (2016) within one order of magnitude. Every other substrate sits above. The chemical synapse spends ten thousand ATP per bit (~10⁻¹⁶ J, five orders above the floor). SRAM and DRAM sit at six to eight orders above. NAND flash at nine. Current generative-model inference, per output bit, sits in the eighteen-to-twenty orders above the floor range. This axis spans the widest dynamic range in the table.

Admissible classes

Citations

Landauer, IBM J. Res. Dev. 5 (1961). · Bérut, Petrosyan, Ciliberto, Nature 483 (2012). · Hong, Lambson, Dhuey, Bokor, Sci. Adv. 2 (2016). · Laughlin, de Ruyter, Anderson, Nat. Neurosci. 1 (1998).

05 · COST TO HOLD

joules per bit per second of retention

Hold energy

The energy spent to keep a bit between writes and reads. Distinct from write cost because the underlying physics is different: holding is a battle with leakage, decoherence, biochemical turnover, or social forgetting. SRAM pays leakage continuously. DRAM pays refresh every sixty-four milliseconds. Flash pays nearly nothing once written. Synaptic structures pay protein-synthesis costs to maintain late-phase LTP. Trapped-ion qubits pay every microsecond of dynamical decoupling. Songlines pay ceremony. The axis collapses the substrate's choice of how to fight entropy at rest.

Admissible classes

Citations

Dennard, IEEE JSSC 9 (1974) — leakage scaling. · Attwell & Laughlin, JCBFM 21 (2001) — neural energy budget. · Frey & Morris, Nature 385 (1997) — protein synthesis in L-LTP.

06 · COST TO RECALL

joules per bit returned on read

Recall energy

The energy dissipated to bring a stored bit back. In well-designed classical memory, recall is effectively free relative to write — SRAM and DRAM read at sense-amp cost, near the substrate floor. In biological memory, recall is essentially free of new transmission cost but is a write event on the trace itself (reconsolidation). In current generative AI, recall is the dominant cost: a generative model reconstructs each output bit by running attention across the full context, paying joules per bit several orders above what biological recall costs. The asymmetry between write and recall, plotted across substrates, is where the design space partitions cleanly.

Admissible classes

Citations

Nader, Schafe, LeDoux, Nature 406 (2000) — recall as write. · Forrester, MIT Whirlwind Project (1953) — destructive read. · Niu et al., TokenPowerBench, AAAI 2026 — measured LLM J/token.

07 · HOW IT IS REACHED

the mechanism that returns a trace

Access regime

How a stored trace is reached. Indexed memory takes an address and returns content (the von Neumann default). Associative memory takes a cue and returns content via overlap. Content-addressable memory takes the content itself and returns a location — Hopfield, Kanerva's sparse distributed memory, hyperdimensional computing. Similarity access takes a vector and returns the nearest neighbor — the ANN-vector-database move. Context-reinstatement, Tulving and Thomson's encoding specificity, requires reproducing the encoding state to recover the trace — how hippocampal recall actually works. Generative-reconstructive access does not retrieve at all; it synthesizes the trace from a schema and a cue — Bartlett's finding in 1932, and what current LLMs do.

Admissible classes

Citations

Hopfield, PNAS 79 (1982). · Kanerva, Sparse Distributed Memory, MIT Press (1988). · Tulving & Thomson, Psychol. Rev. 80 (1973). · Bartlett, Remembering, Cambridge (1932).

08 · HOW IT EVOLVES WHILE HELD

how the trace changes between write and recall

Fidelity dynamics

How the trace evolves while held — distinct from how long it lives. Stable traces do not change. Refresh-dependent traces persist only while the refresh process runs. Decaying traces follow a published forgetting curve — Ebbinghaus's exponential, refined by Murre and Dros to a power law with a twenty-four-hour upward boost from sleep consolidation. Reconsolidating traces are rewritten on every recall. Interference-limited traces degrade as new traces overlap them (Hopfield crosstalk, proactive and retroactive interference). Generative-reconstructive traces are not stored at all between recalls — they are re-synthesized each time.

Admissible classes

Citations

Ebbinghaus, Über das Gedächtnis (1885). · Murre & Dros, PLOS ONE 10 (2015). · Nader et al., Nature 406 (2000). · Bartlett, Remembering (1932).

09 · THE FORM OF THE TRACE

how the trace is etched into the substrate

Representational form

The form a trace physically takes. Symbolic: discrete tokens, addresses, B-tree keys, syllables, knots. Distributed: weights spread across many elements, individually meaningless, Hopfield-style. Hyperdimensional: Kanerva's high-dim binary or bipolar vectors, with O(1) similarity. Parametric: compiled into adjustable values — LTP, neural-network weights, gradient-encoded. Episodic: indexed by context and time, sparse and pattern-separated. Procedural: compiled into action patterns, cerebellar or striatal. Each form has different capacity-fidelity-energy trade-offs, all measured in the literature.

Admissible classes

Citations

Hopfield, PNAS 79 (1982). · Kanerva, Cognitive Computation 1 (2009). · Kandel, Science 294 (2001). · Squire, Psychol. Rev. 99 (1992). · Tulving, in Organization of Memory (1972).

10 · HOW IT IS WRITTEN

how a trace enters memory

Write regime

How a trace enters the substrate. Imperative writes are explicit stores — the von Neumann default. Hebbian writes strengthen connections that co-activate; Donald Hebb's 1949 proposal, the molecular substrate of synaptic plasticity. Gradient writes adjust weights to reduce error, the training-time write of every neural network. Implicit writes are side-effects of computation — the KV cache is the canonical example, built as a byproduct of attention. Consolidation transfers traces from a faster, lossier tier to a slower, more durable one — the CLS framework's central operation, demonstrated in hippocampal-cortical transfer over sleep. Acquisition is the CRISPR move: explicit capture of an environmental signal into a hereditary substrate.

Admissible classes

Citations

Hebb, The Organization of Behavior (1949). · McClelland, McNaughton & O'Reilly, Psychol. Rev. 102 (1995). · Barrangou et al., Science 315 (2007). · Nuñez et al. (Doudna lab), Nat. Struct. Mol. Biol. 22 (2015).

11 · TRUST ON RECALL

what guarantee travels with a recalled trace

Verifiability

The strength of the guarantee that what is recalled is what was stored. None: raw retrieval with no check (DRAM without ECC, an LLM's parametric memory). Parity or error-correcting codes: low-cost statistical assurance, the standard for silicon memory and synthetic DNA. Cryptographic signature: unforgeable provenance, the move every blockchain makes. Source-monitoring: the cognitive analog — Johnson's framework for attributing memories to their origin. Multi-encoder cross-check: the Vedic pātha system's recurring solution and the quipucamayoc's, where independent encodings of the same content must agree. Affinity selection: the immune system's mechanism — only receptors that bind the original antigen well enough are retained.

Admissible classes

Citations

Shannon, BSTJ 28 (1949) — secrecy systems. · Goldwasser, Micali, Rivest, SIAM J. Comput. 17 (1988). · Johnson, Hashtroudi, Lindsay, Psychol. Bull. 114 (1993). · Staal, Nature 322 (1986) — Vedic multi-pātha. · Victora & Nussenzweig, Annu. Rev. Immunol. 30 (2012).