paper
active
2024
paper:entropy-26-00481

Self-Improvising Memory: A Perspective on Memories as Agential, Dynamically Reinterpreting Cognitive Glue

ByMichael Levin
DOI 10.3390/e26060481
Care and CognitionExpanded cognition & sentience frameworksCare-based cognition and selfless AIMultiscale Memory SystemsBioelectric membrane voltage patterns as morphogenetic informationBiosemiotics: the study of meaning-making among and within agentsBowtie architecture: compression during learning; creative reinterpretation during recall and generalizationbrain fills in blind spot via controlled hallucination (inpainting)Self-Improvising Memory frameworkConfabulation: active modification and reinterpretation of memory data to preserve coherent self-model and narrativeevolution of communicationfragmentation of traumatic memories in PTSDMorphogenetic MemoryPattern completion and confabulation in anatomical spacePolycomputationPolycomputation: life as a diverse set of embodied perspectives—nested agents who interpret each other's and their own past messagesSelflet: temporal slice of continuous being, analogous to bread-loaf slices in Special Relativity spacetime modelShip of Theseus of the body+2 more

TL;DR

Memory's function is not fidelity but salience preservation through active reinterpretation—a claim Levin defends by unifying phenomena from caterpillar-to-butterfly metamorphosis (where training-induced memories survive radical brain remodeling and are remapped onto an entirely different body plan), to polyploid newt kidney tubules that maintain normal diameter by substituting a single-cell cytoskeletal-bending mechanism for multi-cell tubulogenesis, to Xenopus tadpoles with ectopic eyes wired to the spinal cord rather than the brain that nonetheless learn effectively in visual assays. The paper introduces the concept of mnemonic improvisation—the dynamic capacity to rewrite and remap information, including memories, onto new substrates and contexts across molecular, physiological, and behavioral scales—and frames it within a bowtie (hourglass) architecture in which compression during encoding necessarily renders the engram increasingly random-looking, making any decoding by a future Self an irreducibly creative, not merely deductive, act. Planarian flatworms, which accumulate somatic mutations through asexual reproduction yet retain extreme regenerative competence and resistance to aging and cancer, are offered as the limit case: their genomes are maximally noisy precisely because evolution committed entirely to large-scale pattern completion over genomic fidelity. The paper argues this implies that morphological and behavioral intelligence arose because of substrate unreliability rather than despite it, generating a positive-feedback competency ratchet, and that memories themselves—as patterns in excitable cognitive media—may possess minimal agency, dissolving the binary between passive data and the computational architectures that hold them, with direct implications for Diverse Intelligence research and the engineering of novel embodied minds.

What to take away

  1. 1. Mnemonic improvisation—the active capacity to rewrite and remap information onto new substrates and contexts—is proposed as the core cognitive mechanism that maintains selfhood across scales from subcellular organelles to evolutionary lineages.
  2. 2. Training-induced associative memories in caterpillars (e.g., conditioned aversion cued by a bright red disk) survive metamorphosis into butterflies despite near-complete brain remodeling, but are remapped so that only high-level salience ('food source') transfers, not body-specific actuator details.
  3. 3. Polyploid newt embryos maintain normal kidney tubule diameter across ploidy levels by adjusting cell number and shape, and when cells are made truly enormous, a single cell substitutes cytoskeletal bending for multi-cell tubulogenesis to accomplish the same morphogenetic goal.
  4. 4. Xenopus tadpoles with ectopic eyes grafted to their tails, connecting to the spinal cord or gut rather than the brain, nonetheless learn effectively in light-mediated visual assays, requiring no additional generations of mutation and selection to integrate the novel sensorimotor architecture.
  5. 5. Injection of an odorant molecule into a frog egg causes the resulting adult to seek that odor during food-foraging, demonstrating remapping from a single-cell biochemical signal across scales to nervous-system-level behavioral output without conventional neural encoding.
  6. 6. In David Glanzman's Aplysia experiments, RNA extracted from trained donors and injected non-specifically into the general vicinity of naïve recipients' nervous tissue induces sensitization memory, indicating that engram decoding is substrate-tolerant and context-reconstructive rather than position-dependent.
  7. 7. Planarian flatworms—which accumulate somatic mutations through asexual reproduction yet exhibit extreme regeneration, cancer resistance, and longevity—are argued to achieve these capacities because evolution invested entirely in large-scale pattern-completion competency rather than genomic fidelity, exemplifying the 'competency ratchet' hypothesis.
  8. 8. The bowtie (hourglass) architecture—a low-dimensional compressed hub through which information must pass before context-sensitive re-expansion—is proposed as a unifying motif across biochemical signaling pathways, bioelectrical circuits, biomechanical networks, autoencoders in machine learning, language, and the egg-to-organism developmental compression cycle.
  9. 9. An open hypothesis is raised as to whether memories themselves, as metastable patterns in excitable cognitive media, possess minimal agency—actively facilitating their own remapping and thereby blurring the binary distinction between passive data and computational architecture, in the spirit of William James's claim that 'thoughts are thinkers.'
  10. 10. As a replicable methodology, the paper proposes moving pattern memories from two-headed planaria to wild-type hosts via tissue implants, and transferring behavioral memories (such as nicotine addiction) from human donors through Anthrobots into rat hosts, as in vivo model systems for testing whether remapped engrams alter recipient behavior (e.g., self-medication).

Peer brief — for seminar discussion

Michael Levin's perspective piece argues that the canonical framing of memory research—focused on substrate, encoding fidelity, and reliable readout—systematically neglects the complementary and arguably more fundamental problem: how agents dynamically reinterpret and remap stored information to preserve salience rather than detail across radical changes in body, environment, and Self. To make this case, the paper ranges across developmental biology, evolutionary biology, neuroscience, and synthetic bioengineering, treating each domain as an instance of the same underlying architecture. The load-bearing finding is that biological systems at every scale implement what Levin terms mnemonic improvisation—active, creative rewriting of information onto new substrates and contexts. Three substrate-level demonstrations anchor the claim. First, caterpillar-to-butterfly metamorphosis retains training-induced associative memories despite near-complete brain remodeling, but only at the level of abstracted salience (food-relevant cue) rather than actuator-specific detail. Second, polyploid newt kidney tubules maintain normal diameter across ploidy levels by substituting single-cell cytoskeletal bending for multi-cell tubulogenesis when cells become sufficiently large—a case where the morphogenetic 'memory' of correct tubule geometry overrides the molecular mechanism used to achieve it. Third, Xenopus tadpoles with ectopic eyes wired to the spinal cord rather than the brain learn effectively in light-mediated visual assays without additional selection, and injection of an odorant into a frog egg shapes adult food-seeking behavior, demonstrating cross-scale remapping from single-cell biochemistry to nervous-system behavior. Planarian flatworms—which accumulate somatic mutations through asexual reproduction yet retain extreme regenerative competence and cancer resistance—are offered as the evolutionary limit case, where selection pressure fell entirely on large-scale pattern-completion competency rather than genomic fidelity. The unifying framework Levin introduces is the bowtie (hourglass) architecture, a compression bottleneck through which complex states are squeezed into a low-dimensional hub before context-sensitive re-expansion. This motif appears in biochemical, bioelectrical, and biomechanical signaling pathways, in machine-learning autoencoders, in the egg-to-organism developmental cycle, in language, and in the engram-to-recall cycle. Because compression removes correlations and renders engrams increasingly random-looking, decoding by a future Self is necessarily creative rather than deductive—an alternative method the paper could have used to formalize this is reservoir computing, which Levin explicitly acknowledges as a candidate tool for modeling how rich biological structures are interpreted on the fly. The paper's central prediction is a competency ratchet: substrate unreliability hides genomic variation from selection, shifting evolutionary pressure onto interpretation and remapping mechanisms, which in turn enables further substrate degeneracy, producing a positive feedback loop that generated intelligence. A secondary speculative hypothesis is that memories, as metastable patterns in excitable media, may possess minimal agency—actively shaping the cognitive systems that hold them and dissolving the sharp boundary between data and thinker. A critical reader would push back most forcefully on the evidentiary standard applied across scales. The paper treats phenomena as homologous instances of mnemonic improvisation—from Aplysia RNA transfer experiments, to planarian bioelectric regeneration, to human confabulation in split-brain patients, to cross-embryo calcium/ATP morphogenetic signaling—but the mechanistic connective tissue between these levels is largely absent. The bowtie architecture is a useful descriptive schema, but asserting that developmental compression, memory consolidation, and language communication share the same underlying computational principle without a formal model risks being unfalsifiable. Whether the 'salience-preserving reinterpretation' observed in metamorphic memory retention is genuinely the same process as morphogenetic pattern completion in polyploid newts, rather than a superficial analogy, is precisely what requires the computational and in vivo experiments Levin defers to future work.

Methods (1)

Frameworks (2)

Findings (20)

Claims (27)

Hypotheses (5)

Questions (3)

Related work— refs + corpus + external arXiv

Cited / in-corpus / arXiv badges show which signals surfaced each row. Multi-source rows weighted higher.

+27 more

Similar preprints — Semantic Scholar

Cited by (1)

Cross-corpus bridges (10)

same_concept_as · Nomic cosine

External markdown files that talk about the same concept as this entity.

  • aboutblank_kb
    How can memories remain and be adaptively used across drastic morphological changes like metamorphosis, where the body architecture and sensory systems are fundamentally reorganized?questions/how-can-memories-remain-and-be-adaptively-used.md0.836
  • aboutblank_kb
    How does biological memory persist across metamorphosis and major structural reorganization when neural architecture is destroyed?questions/how-does-biological-memory-persist-across-metamorphosis-and.md0.831
  • aboutblank_kb
    How can memories persist and remain adaptive across radical refactoring of brain tissue during metamorphosis?questions/how-can-memories-persist-and-remain-adaptive-across.md0.821
  • aboutblank_kb
    How do molecular events expand into complex anatomical structures with persistent pattern memories?questions/how-do-molecular-events-expand-into-complex-anatomical.md0.818
  • aboutblank_kb
    What mechanisms preserve memories during radical morphological changes like metamorphosis?questions/what-mechanisms-preserve-memories-during-radical-morphological-changes.md0.811
  • aboutblank_kb
    How can organisms with extremely variable or messy genomes (like planaria) exhibit superior morphological control compared to organisms with stable genomes?questions/how-can-organisms-with-extremely-variable-or-messy.md0.810
  • aboutblank_kb
    Can memories function as active agents that help reify and maintain the self through circular causality?questions/can-memories-function-as-active-agents-that-help.md0.793
  • aboutblank_kb
    How do thoughts and memories scale up from fleeting patterns to persistent thinkers capable of generating novel thoughts?questions/how-do-thoughts-and-memories-scale-up-from.md0.791
  • aboutblank_kb
    What role do memory and forgetting play in organisms' ability to evolve and persist?questions/what-role-do-memory-and-forgetting-play-in.md0.787
  • alexander
    Source: https://studio.ribbonfarm.com/p/beyond-hyperanthropomorphismarticles/essay-site/2022-12-23_Stefan-Lesser_studio.ribbonfarm.com_p-beyond-hyperanthropomorphism_64d412.md0.753