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

ByMichael Levin

TL;DR

Memory systems in biological organisms preserve salience rather than fidelity, and this reinterpretive commitment—termed Mnemonic Improvisation by Michael Levin—is the generative source of morphological and behavioral intelligence rather than an architectural flaw. Published in Entropy 2024, volume 26, article 481, the argument is anchored in three converging substrate references: metamorphosing insects that retain associationally generalized (not detail-specific) engrams across near-total brain remodeling; polyploid newts (engineered at multiple ploidy levels) whose kidney tubules maintain normal diameter by substituting cytoskeletal bending for cell–cell communication when cell number drops to one; and Xenopus tadpoles bearing ectopic eyes connected to spinal cord or gut rather than brain, which nonetheless learn visual discriminations effectively. The Bowtie Architecture—a compression bottleneck forcing low-dimensional encoding followed by context-sensitive re-inflation, instantiated in developmental bioelectric signaling, the egg-to-organism reproductive cycle, and machine-learning autoencoders—is proposed as the universal implementation mechanism. Engrams treated as stigmergic prompts passed between temporally separated Selflets, combined with polycomputing (the same physical substrate simultaneously supporting multiple computational interpretations for different observers), explains why Glanzman-style RNA extract injections into the general vicinity of Aplysia nervous tissue suffice to transfer behavioral sensitization without precise placement. The paper argues this implies that intelligence—morphogenetic and behavioral—arose because of substrate epistemic vulnerability, not despite it, and that the same principles are actionable for engineering genuinely bio-inspired AI architectures and regenerative biomedical interventions that target how cells interpret, not merely receive, biochemical and bioelectrical stimuli.

What to take away

1. Levin proposes that biological memory systems optimize for salience preservation rather than fidelity, a principle he terms Mnemonic Improvisation, which operates across scales from intracellular engrams to evolutionary lineages.
2. Polyploid newts engineered at multiple chromosome copy numbers complete normal development, with kidney tubule diameters held constant by reducing cell number and, at extreme ploidy, by a single cell bending via cytoskeletal mechanisms instead of cell–cell tubulogenesis signaling—demonstrating that morphogenetic goals override molecular implementation details.
3. Xenopus tadpoles bearing ectopic eyes connected to the spinal cord or gut rather than the brain successfully perform light-mediated learning tasks, showing that sensorimotor remapping to non-canonical neural substrates requires no multigenerational selection.
4. In Glanzman's Aplysia experiments, RNA extract from trained animals injected non-specifically into the general vicinity of naïve nervous tissue transfers long-term sensitization, indicating that recipient neural tissue extracts behavioral meaning without requiring precise molecular placement.
5. The Bowtie Architecture—a low-dimensional compression bottleneck flanked by intelligent encoding and context-sensitive decoding agents—is identified as the shared structural motif across developmental morphogenesis, the egg-to-organism reproductive cycle, autoencoder machine learning models, and inter-human language communication.
6. Injecting an odorant molecule into a single frog egg cell results in adult animals that preferentially seek that odor during food-finding behavior, demonstrating cross-scale remapping from a biochemical signal in cytoplasm to a modified genetically encoded nervous system and behavioral output.
7. Levin raises the open hypothesis that memories, as metastable patterns in excitable cognitive media, may themselves possess basal agency—actively facilitating their own remapping and survival across architectures—instantiating William James's claim that 'thoughts are thinkers' as a testable biological proposition.
8. Planarian flatworms, which have the noisiest (mixoploid) genomes among studied organisms, also exhibit the strongest regeneration, immortality, and cancer resistance, a century-old puzzle Levin resolves by arguing that Constructive Neutral Evolution placed all selective pressure on competency mechanisms rather than structural genome fidelity.
9. To test memory remapping empirically, Levin proposes transferring pattern memories from two-headed planarians to wild-type hosts via tissue implants and transferring behavioral memories such as nicotine addiction from human donors through Anthrobots into rat hosts to assess whether recipients self-medicate—a replicable experimental design using existing biobot technology.
10. The paper hypothesizes that consciousness specifically may be palpated uncertainty about one's own memories and internal states, extending Solms's proposal that consciousness is uncertainty about the external world to include interoceptive and mnemonic uncertainty as primary drivers of subjective experience.

Peer brief — for seminar discussion

Levin's perspective piece, published in Entropy 2024 (vol. 26, no. 481), argues that the dominant framing of memory research—focused on substrate fidelity and reliable read-out—misses the functionally more important dimension: the active, creative reinterpretation of stored information to preserve salience as bodies, environments, and cognitive architectures change. The paper introduces Mnemonic Improvisation as the name for this capacity, defined as the dynamic ability to remap information onto new media and new contexts across behavioral, genetic, and physiological levels. An alternative framing Levin could have used but explicitly sets aside is standard error-correction or de-noising, which targets low-level detail reconstruction rather than high-level meaning extraction. The load-bearing empirical anchors are drawn from four domains. First, memory retention across metamorphosis: caterpillar-to-butterfly transitions involve near-total brain remodeling, yet associative memories survive in a form remapped from leaf-seeking actuator commands to generalized food-valence signals usable in a radically different body. Second, polyploid newt kidney tubules maintain normal diameter across multiple engineered ploidy levels by substituting cytoskeletal bending for cell–cell tubulogenesis when cell number is reduced to one, illustrating that goal-state homeostasis overrides molecular-implementation fidelity. Third, Xenopus tadpoles bearing ectopic eyes wired to spinal cord or gut tissue rather than brain learn visual discriminations successfully, requiring no generations of selection. Fourth, Glanzman's Aplysia RNA-extract experiments demonstrate behavioral memory transfer via non-specific injection, implying that recipient tissue extracts meaningful engrams without precise molecular positioning. Levin unifies these phenomena under the Bowtie Architecture—a low-dimensional compression bottleneck flanked by intelligent encoding and context-sensitive decoding—instantiated in developmental morphogenesis, the egg-to-organism cycle, autoencoder neural networks, and language-mediated human communication. The central implication is an intelligence ratchet: because biological substrate is epistemically vulnerable (subject to mutation, aging, hacking by other organisms, and developmental remodeling), selection converged on competency mechanisms that actively reinterpret information rather than defend its literal details, and behavioral and morphological intelligence arose because of this vulnerability rather than in spite of it. Levin further speculates that memories, as metastable excitable-medium patterns, may constitute minimal agents in their own right—a continuum from passive data through active algorithms to basal-cognitive thought-forms—and proposes testing this by measuring Integrated Information in gene regulatory networks before and after associative memory training. One prediction named explicitly is that planarian pattern memories can be transferred to wild-type hosts via tissue implants, with preliminary unpublished data cited. A critical reader would push back most forcefully on the methodological conflation of phenomena across wildly different scales and substrates: the paper treats RNA-extract memory transfer in Aplysia, ectopic eye learning in Xenopus, polyploid newt morphogenesis, and human confabulation as instances of the same underlying Mnemonic Improvisation mechanism, but no common mechanistic account is provided, and each phenomenon has domain-specific alternative explanations. The Bowtie Architecture is a structural analogy, not a demonstrated causal mechanism, and invoking it across developmental biology, machine learning, and human language risks substituting a compelling metaphor for a testable process model.

Findings (5)

Commissurotomy (Split-Brain) Confabulation
Split-brain patients whose left hand performs unexpected movements generate narrative explanations to preserve self-model; demonstrates active memory reinterpretation.
David Glanzman's RNA Transfer Experiments
Demonstration that trained sea slug (Aplysia) brain extracts transfer behavioral memories to naïve animals; shows memory substrate-independent remapping.
Polyploid Newt Kidney Tubule Adaptation
Engineered polyploid newts with abnormally large cells and extra chromosomes develop normal kidney tubule diameters by switching molecular mechanisms on-the-fly.
Memory Transfer Across Metamorphosis
Empirical observation that caterpillars retain learned memories through pupation despite radical brain refactoring, suggesting memory as salience rather than fidelity.
Ectopic Eye on Tadpole Tail
Tadpoles bearing eyes placed on tails can see and learn effectively despite novel neural connections (to spinal cord or gut), demonstrating plastic sensorimotor reinterpretation.

Claims (2)

Liquid Brain Hypothesis
Ant colonies are 'liquid brains' due to shifting connectivity; extended metaphorically to all brains being 'liquid in time' through reinterpreting mental content across moments.
Stigmergic Memory Model
Memories are stigmergic notes left by past versions of self; communication between temporal slices of the same agent analogous to lateral inter-agent communication.

Hypotheses (3)