Iterated Insights

How Iterative Updating Produces Descent with Modification in Long-Term Memory

Jared Edward Reser PhD

Abstract

The iterative updating model describes thought as a succession of partially overlapping working memory states. At each moment, some active representations are retained, some subside, and others are added through multiassociative search. Coactive representations also modify the associative structure of long-term memory, changing which representations will be retrieved together in the future. This article extends that model from the short timescale of individual thoughts to the long timescale of intellectual development. It proposes that ideas form historical lineages as they are repeatedly retrieved, combined with new information, differentiated by context, and reconsolidated in modified form. An idea is therefore treated as neither a fixed sentence nor an immutable neural object, but as a family of causally connected representational versions.

This framework introduces the concepts of idea tokens, idea versions, conceptual lineages, trace-spanning continuity, conceptual differentiation, conceptual speciation, and conceptual synapomorphies. It argues that the resulting history is usually better represented as a temporally directed phylogenetic network than as a simple tree because ideas regularly inherit from multiple antecedents, incorporate information acquired from other minds, converge independently on similar solutions, and reactivate dormant branches preserved in memory or external records. Timestamped notes, drafts, diagrams, publications, and conversations can serve as partial observations of this history. When combined with semantic analysis, explicit source information, chronological constraints, and human validation, they could support the reconstruction of an individual’s intellectual development.

The framework generates empirical predictions for longitudinal behavioral and neuroimaging studies and suggests an artificial intelligence architecture that records the provenance of its own conceptual changes. The central claim is that the iterative updating of working memory supplies local transformations, learning preserves selected products, and repeated retrieval organizes those products into branching and recombining lineages across a lifetime.

Keywords: associative memory, cognitive development, concept learning, conceptual differentiation, cultural evolution, long-term memory, memory reconsolidation, multiassociative search, phylogenetic network, working memory

Introduction

A phylogenetic tree depicts present forms as descendants of earlier forms. Branches preserve historical continuity while also recording divergence. A tetrapod lineage does not begin anew with every generation. Existing organization is retained, altered, and transmitted. Over long periods, accumulated modifications produce distinguishable groups whose common ancestry can still be reconstructed.

Thought may have an analogous history. Memories, concepts, explanations, and theories do not ordinarily appear in final form. They develop as earlier representations are retrieved into working memory, combined with other representations, interpreted in new contexts, and returned to long-term memory with altered associative relationships. Some variants disappear. Others are rehearsed, elaborated, generalized, or divided into more specialized forms. Separate lines of thought may later merge. An idea encountered in a book or conversation may enter an existing lineage and redirect it. A forgotten formulation may survive in a notebook and become active again years later.

The iterative updating model provides a mechanistic starting point for explaining this process (Reser, 2016, 2024). In that model, working memory is updated through partial replacement. Some representations remain active from one state to the next while other representations enter and leave. The retained contents supply continuity and act together as parameters for multiassociative search. The result of one search becomes part of the conditions governing the next search. Thought therefore advances through a succession of overlapping and interdependent states rather than a sequence of isolated mental frames.

The same episodes that generate thought can also modify long-term memory. Coactive representations strengthen or weaken relationships among their underlying neural assemblies, and new combinations become more available during later retrieval. When these changes persist, the products of one thought episode influence thoughts occurring much later. This makes it possible to extend the iterative updating model beyond the microgenesis of a single thought. The present article develops that extension as a phylogeny of ideas.

The word phylogeny is normally reserved for evolutionary relationships among organisms, populations, or genes. Conceptual ontogeny or idea genealogy may be more literal terms for development within one person. Phylogeny remains useful because the proposed history contains descent with modification, branching, differential retention, convergence, and recombination. The analogy becomes scientifically productive when its units and limits are defined rather than assumed.

Iterative Updating as the Microgenesis of Thought

Working memory temporarily maintains a limited collection of representations for use in ongoing processing. In embedded-process accounts, the contents of working memory are activated portions of long-term memory rather than copies transferred into a separate storage device (Cowan, 1988). The iterative updating model adds a temporal organization to this relationship. Persistent neural activity causes neighboring working memory states to overlap (Fuster, 2009; Miller & Cohen, 2001). Short-lived synaptic changes may also preserve latent contents without uninterrupted firing (Mongillo et al., 2008; Rose et al., 2016). Some representations remain available while others are replaced, so the current state is a modified continuation of the preceding state (Reser, 2016, 2024).

Each active representation consists of a distributed and context-sensitive pattern. Its neural composition need not be identical every time it is activated. What a concept contributes to processing depends partly on the other representations that are active with it. The concept of iteration, for example, will recruit different features when considered alongside working memory, evolution, computer programming, or visual imagery. Its identity remains recognizable even though its active implementation and immediate meaning vary.

The active set conducts a multiassociative search. Excitatory and inhibitory effects from the contents of the focus of attention, the short-term store, sensory systems, episodic memory, procedural systems, and motivational structures converge on possible additions to working memory. The convergence of distributed activations is consistent with models of multiregional retroactivation (Damasio, 1989). A candidate associated with several current contents can receive more combined support than a candidate associated strongly with only one. The selected addition changes the active constellation and redistributes the influence of the representations that remain. This new configuration then searches again.

The model therefore contains a recursive causal cycle. The current state is partly a product of the preceding search and partly the starting condition for the next search. Over several iterations, intermediate results can accumulate, subproblems can be solved and recombined, mental imagery can be progressively modified, and an initially vague problem can become a more explicit solution. The sequence can also be learned. Repeated coactivity alters associative weights, allowing later searches to recover useful transitions more readily (Hebb, 1949; Anderson, 1983; Collins & Loftus, 1975).

This final property opens the path to a long-term theory. A working memory sequence does not merely move through an unchanged store. It can revise the store through which future sequences will move. Every consequential episode therefore has two products: a transient progression of active states and a persistent change in the probability structure of later thought.

From State-Spanning Coactivity to Trace-Spanning Continuity

The original formulation of the model uses state-spanning coactivity to describe representations that remain active across consecutive neural states (Reser, 2016). This mechanism can explain continuity over seconds, but sustained firing cannot connect two episodes separated by weeks or years. Long-interval continuity must be carried by more durable changes in synaptic organization, systems-level consolidation, behavioral habits, and external records (McClelland et al., 1995).

Three nested timescales should therefore be distinguished. Within a thought, persistent activity produces overlap among rapidly successive working memory states. Between nearby episodes, temporary potentiation, episodic context, and recent activation make selected contents easier to reinstate, while temporal context helps organize successive experiences in memory (Howard & Kahana, 2002). Across longer periods, learning preserves altered relationships among representations and changes the attractor structure of the memory network. Later retrieval reactivates some portion of this inherited organization in a new context.

The term trace-spanning continuity can designate this third form. It is continuity by causal inheritance rather than uninterrupted activity. A concept activated today may share no currently firing neurons with an activation of the concept a year ago. Nevertheless, the earlier episode helped shape the synaptic and relational structure from which the later activation is reconstructed. The two are historically connected.

Retrieval is especially important because reactivation can make a memory susceptible to modification. Experimental work on reconsolidation shows that retrieved memories can again depend on plastic processes and can incorporate new information (Nader et al., 2000; Hupbach et al., 2007). Research on overlapping memories likewise shows that new learning can produce either integration or differentiation in hippocampal and prefrontal representations (Zeithamova et al., 2012; Schlichting et al., 2015; Schlichting & Preston, 2015). These findings do not establish a phylogeny of ideas by themselves, but they supply mechanisms through which ancestral content can be reinstated and altered.

Figure 1. Nested timescales in the phylogeny of ideas. Iterative updating transforms working memory over seconds. Learning changes long-term memory after an episode. Repeated retrieval and modification produce conceptual lineages over months and years.

Units of Conceptual Evolution

A theory of conceptual ancestry requires a defensible unit of analysis. Treating each word as an idea would fragment meaningful structures. Treating an entire worldview as one idea would hide the changes that need to be explained. A multilevel vocabulary resolves this problem.

An idea token is a particular activation or expression of an idea. A spoken explanation, a paragraph in a draft, and a neural reinstatement during silent reasoning are different tokens even when they instantiate closely related content. An idea version is the relatively stable organization that several related tokens express during a period of development. An idea lineage is the temporally ordered family of versions connected by causal inheritance. A schema is a larger organized structure that can contain several lineages and guide their coordinated activation.

This distinction separates occurrence from historical identity. Because neural ensembles are distributed and context dependent, no single activation needs to reproduce an earlier activation exactly. A later token belongs to the same lineage when it inherits enough distinctive organization, inferential role, or causal dependence from earlier versions. Identity through time is consequently genealogical. It depends on the pathway of transformation as well as present similarity.

The same principle applies in biology. Descendants can differ substantially from ancestors while remaining part of the same lineage. Conversely, unrelated lineages can evolve similar features under similar conditions. Conceptual reconstruction must therefore distinguish homology, similarity due to shared history, from analogy, similarity produced independently.

Table 1 defines the principal units introduced by this framework.

Table 1. Principal units in the phylogeny of ideas

Term

Definition

Idea token

One activation or external expression of an idea in a particular episode

Idea version

A temporally bounded organization shared by a cluster of related idea tokens

Conceptual lineage

A sequence or network of idea versions connected by descent with modification

Trace spanning continuity

Causal continuity between temporally separated activations carried by persistent memory structure or external records

Conceptual differentiation

Increasing functional or representational separation among descendants of a shared antecedent

Conceptual speciation

Persistent differentiation sufficient for descendants to be retrieved and used as distinct concepts

Conceptual synapomorphy

A distinctive derived feature that appears in one version and is inherited by its descendants

Documentary fossil

An external record that preserves evidence of an earlier idea version

Phylogenetic network

A temporally directed representation allowing branching, multi-parent inheritance, convergence, and fusion

 

Conceptual Differentiation

Ideas often begin as partially differentiated organizations. An early formulation may combine several intuitions without distinguishing their mechanisms, domains, or implications. Iterative analysis exposes the representation to different constellations of associated material. Features that predict successfully in one context may become strengthened there, while other features become more strongly connected to another context. Repeated contextual specialization can divide one broad representation into separately retrievable descendants.

This process can be described as conceptual differentiation. Suppose a person begins with a general intuition that successive brain states overlap. When this intuition is considered alongside neurophysiology, it develops into an account of staggered persistent activity. When considered alongside working memory, it develops into iterative updating. When considered alongside associative retrieval, it develops into multiassociative search. When considered alongside mental imagery, it develops into progressive imagery modification. When considered alongside artificial intelligence, it becomes a proposed computational architecture. These descendants share an identifiable origin but acquire different inferential roles.

Conceptual speciation occurs when differentiation becomes sufficiently stable that the descendants function independently. Several criteria can operationalize this threshold. The descendant concepts should be elicited by different contexts, support different predictions, participate in different associations, and remain independently available without being treated as interchangeable. Their neural or computational representations should also show reliable separation while retaining measurable similarity to the inferred ancestor.

Prediction error can accelerate differentiation. If one undivided concept produces conflicting expectations in two environments, separating it into context-specific variants reduces interference. Inhibition can also contribute by suppressing an inappropriate interpretation while allowing an alternative to strengthen. Over time, a broad category may divide into subcategories because the division improves prediction and action.

Differentiation should not be equated with inevitable progress or increasing complexity in every descendant. Evolution can eliminate features as well as add them. Expertise often produces both differentiation and compression. An expert distinguishes cases a novice treats as equivalent, yet the expert may also execute a once laborious reasoning sequence through a compact schema. The total knowledge system becomes more articulated even when individual operations become shorter and more automatic.

An Evolutionary Grammar Already Present in the Model

The iterative updating model already depicts most of the transformations needed for a long-term conceptual phylogeny. Its figures were originally intended to describe thought over seconds, but several have direct long-timescale counterparts.

Table 2. Evolutionary interpretations of operations in the iterative updating model

Operation in the iterative updating model

Interpretation in a conceptual phylogeny

Merging separate subsolutions in Figure 31

Recombination of previously distinct conceptual lineages

Revisiting an intermediate state and altering it in Figure 34

Branching from a shared ancestral formulation

Linking the beginning of a sequence directly to its endpoint in Figure 38

Compression of a repeatedly traversed lineage into a learned shortcut

Employing a previously learned schema in Figure 39

Inheritance of an organized higher order structure

Transfer learning in Figure 40

Reuse or exaptation of an established lineage in a new domain

Iterative inhibition in Figure 41

Rejection of candidate variants during selection

Reconciling separate situations with the same concept in Figure 43

Convergence on a similar solution from different starting points

Multiassociative learning in Section 4.2

Persistent modification of the substrate governing future descendants

 

Figure 34 is particularly revealing. A line of thought returns to an intermediate state and continues along a different route. The original and revised sequences share an ancestor and then diverge. At the scale of a single reasoning episode, this is a fork in a search trajectory. If both outcomes are stored and later elaborated separately, the fork becomes a durable branch in conceptual memory.

Figure 31 supplies the complementary operation. Two independently developed subsolutions are reinstated and combined into a hybrid state. This resembles reticulate evolution, in which a descendant inherits from more than one prior lineage. Figures 38 and 39 show how extended sequences become compressed into schemas that can guide subsequent thought without reproducing every intermediate state. Figure 40 then shows inherited organization being applied outside its original setting. These mechanisms together form a microevolutionary grammar of ideas.

Variation Selection and Retention

An evolutionary account requires an explanation of variation and differential retention. Multiassociative search generates variation because a familiar set of representations can converge on different additions under different contexts, goals, physiological states, and histories of learning. The process is constrained rather than random. Existing associations determine the field of plausible candidates, while novelty, uncertainty, emotion, and expected reward alter their priorities.

Selection occurs at several points. Candidate representations first compete for entry into the focus of attention. Once active, they differ in how long they remain available and how extensively they are elaborated. Thought sequences differ in whether they are encoded, rehearsed, or connected to reward and prediction error. Consolidated structures subsequently differ in how often they are retrieved and whether they continue to organize successful behavior. Social communication adds another filter because ideas that are expressed, understood, and reused can enter other cognitive systems.

This framework resembles Campbell’s principle of blind variation and selective retention, but it does not require cognitive variation to be blind (Campbell, 1960). Human thought uses goals, learned search strategies, and explicit evaluation. Much conceptual change is directed. The evolutionary feature lies in the production of alternatives followed by unequal persistence, not in a claim that all alternatives arise randomly.

Retention also should not be confused with truth. Repetition, emotional salience, social reinforcement, identity protection, and ease of retrieval can preserve inaccurate ideas. A phylogeny records what survived and what descended from what. Epistemic evaluation requires a separate analysis of evidence, prediction, and correspondence with the world.

Why Idea Histories Form Networks

A conventional tree permits one lineage to split into descendants but does not allow separated branches to reunite. Conceptual development repeatedly violates this restriction. A new theory can inherit a mechanism from one research tradition, a formal tool from another, and an analogy from a third. Reading and conversation introduce structures developed in other minds. Old and new memories become coactive. Multiassociative search is therefore multi-parental by design.

Cultural phylogenetic research has shown that tree methods can recover useful historical structure from languages and narratives (Gray & Atkinson, 2003; Tehrani, 2013). It has also confronted horizontal transmission, borrowing, and mixture (Greenhill et al., 2009). These complications are even more pronounced within a thinking individual because integration among lineages is a normal cognitive operation rather than an exception.

The appropriate data structure is a temporally directed phylogenetic network. Each node represents a dated idea version. Directed edges represent inherited influence. A node may have several parents when a new formulation integrates multiple antecedents. Apparent cycles can be avoided by treating every reactivation as a new version: a later idea can return to an earlier formulation, but the later reinstatement remains a new event with a later timestamp.

Figure 2. A phylogenetic network of ideas. An ancestral formulation differentiates into two branches. Descendants from the branches later combine with an external source to produce a multi-parent synthesis. A dormant branch can also be reactivated by an archived record.

Tree views can still be useful as simplified projections. They may display the dominant parent of each concept or isolate one mechanism’s history. The underlying record should preserve reticulation so that a clean visualization does not erase the actual sources of an idea.

External Records as Fossils and Propagules

Human memory is supported by external symbolic storage. Notes, diagrams, drafts, recordings, publications, and correspondence preserve products of earlier cognitive states after their active neural traces have changed. Such records can function as documentary fossils because they supply observable evidence about prior idea versions.

The fossil analogy is incomplete in an instructive way. A biological fossil normally records a form without restoring it to the reproducing population. An intellectual record can be reread. Its structure can reenter working memory, reactivate associations, and participate in new learning. A note is therefore both a fossil and a propagule. It preserves an earlier form and can seed a future descendant.

External records also stabilize lineages against continual reconstructive drift. An internally remembered argument may gradually change without its author recognizing the change, consistent with the constructive character of remembering (Bartlett, 1932). A dated draft provides a fixed comparison. Differences between drafts reveal additions, deletions, terminological shifts, changes in causal structure, and branch points that autobiographical memory alone may miss.

Documents remain partial observations. A paragraph records an expressed product, not the entire working memory state that generated it. Authors omit intermediate reasoning, privately reject alternatives, and revise prose for reasons unrelated to conceptual change. The documentary history must therefore be treated as a sampled fossil record rather than a transparent transcript of long-term memory.

Reconstructing a Cognitive Phylogeny

A practical reconstruction can combine documentary evidence with computational analysis and author validation. The procedure should begin by assembling a chronologically ordered corpus. Relevant sources include notebooks, manuscript versions, diagrams, blog posts, emails, recorded explanations, search histories, and conversation transcripts. Original timestamps and source relationships should be preserved.

The corpus can then be decomposed into analyzable characters. These may include propositions, definitions, causal claims, distinctions, analogies, terms, diagrams, predictions, and applications. Each appearance becomes an idea token. Similar tokens can be grouped into candidate versions, but the grouping should retain their dates and source locations.

Lineage inference must use more than semantic similarity. A later statement is more likely to descend from an earlier one when it preserves unusual terminology, a distinctive relational structure, an uncommon example, or a specific diagrammatic organization. Such features can be treated as conceptual synapomorphies. Explicit self citation, links between files, revision histories, and statements about how an idea developed provide additional evidence.

Chronology constrains the direction of ancestry. A later document cannot be the parent of an earlier one, although both may descend from an undocumented precursor. When no surviving record captures a necessary intermediate step, the reconstruction may posit a latent ancestor and assign it an uncertainty estimate. Similarity without an evidential pathway should be labeled convergence rather than assumed inheritance.

The resulting graph should use typed edges. Useful edge classes include retention, elaboration, differentiation, contradiction, compression, transfer, integration, revival, and rejection. Each edge should carry a confidence score derived from temporal proximity, shared derived features, explicit provenance, contextual evidence, and author confirmation. The graph can then be inspected for roots, branch points, fusion nodes, dormant intervals, bursts of development, and repeated cycles of compression and re elaboration.

The development of the iterative updating model itself offers a candidate case study. The 2016 account of incremental change in state-spanning coactivity provides an identifiable ancestral formulation. Later work differentiates this account into a model of working memory updating, multiassociative search, progressive modification, schema use, transfer learning, and machine consciousness (Reser, 2016, 2024). Earlier manuscripts, conference materials, website posts, and dated diagrams could reveal when these branches first appeared and which later formulations recombined them. The present proposal would itself become a new descendant produced by integrating iterative updating with evolutionary reconstruction.

A reconstructed network should be evaluated against known history. A portion of the corpus can be withheld while the model infers missing links or dates. The author can independently report remembered influences and branch points before seeing the reconstruction. Agreement among document history, computational inference, and prospective self reports would provide stronger evidence than any one source alone.

A Neurocognitive Research Program

Current neuroimaging cannot recover decades of idea ancestry from a single scan. Distributed representations do not contain readily readable dates, and several histories can produce similar present states. A prospective longitudinal design is more tractable.

Participants could learn a complex and initially unfamiliar theoretical domain over several months. At regular intervals they would define central concepts, draw concept maps, judge similarities, solve transfer problems, and explain their reasoning aloud. These behavioral products would provide dated samples of conceptual organization. Periodic fMRI sessions could measure multivoxel patterns during controlled retrieval, while EEG or MEG could estimate the temporal order in which related representations are reinstated. High-frequency single-participant designs have already demonstrated that repeated neuroimaging can track changes within one person across days and months (Triana et al., 2024).

The study could deliberately create opportunities for differentiation and integration. One broad concept would first be taught in a common context and later applied in two environments with different predictive requirements. Another condition would teach two separate structures that can later be combined to support a novel inference. This design would allow the following predictions to be tested.

1. Descendant representations should retain more neural and semantic similarity to their documented ancestors than to matched control concepts.

2. When one representation differentiates into context-specific descendants, cross-context substitutability should decline while within-context decoding becomes more reliable.

3. A synthesis should be preceded by reinstatement of information from both parent lineages. The strength of joint reinstatement should predict successful inference.

4. Repeated traversal of a reasoning sequence should eventually permit a more direct transition from initial conditions to the solution. Intermediate states should become less behaviorally necessary, corresponding to the compression depicted in Figure 38 of the iterative updating model.

5. Variants receiving more elaboration, reward, successful application, or subsequent retrieval should show greater long-term accessibility.

6. Reintroducing an archived early formulation should selectively reactivate its descendants and may produce a new branch that differs from the branch produced by unaided recollection.

Representational similarity analysis can test inherited structure across sessions. Pattern separation and integration can be assessed by comparing changes in neural geometry (Schlichting et al., 2015). Time resolved methods can test whether parent representations precede a new synthesis. Textual and diagrammatic measures can be compared with neural measures without assuming that either provides a complete description of the concept.

Formal Representation

Let M(e) denote the long-term memory network before cognitive episode e. Let X(e,k) denote the active working memory configuration at iterative step k within that episode. Let S(e,k) represent current sensory or informational input and G(e) represent current goals. The within-episode transition can be written as:

X(e,k+1) = U[X(e,k), M(e), S(e,k), G(e)]

The function U includes retention, subtraction, addition, contextual reweighting, and multiassociative search. It produces the next working memory state from the current state and the memory network in which the search occurs.

After the episode, learning changes the memory network:

M(e+1) = M(e) + eta(e)L[X(e,0:K), O(e)]

Here, L represents plastic change induced by the episode, O(e) represents relevant outcomes such as prediction error and reward, and eta(e) represents an effective learning rate. The second equation changes the conditions under which future applications of U will operate. The full theory therefore consists of a fast iterative process nested within a slower self-modifying process.

Let C(i,e) represent an idea version realized within M(e). A proposed ancestry relation from C(i,e) to C(j,f), where e is earlier than f, should satisfy both inheritance and influence. The later version should preserve identifiable structural or functional features of the earlier version, and the earlier version should have contributed causally to the later one’s formation. In an artificial system this contribution can be tested through ablation. In a human history it must usually be inferred from documents, retrieval reports, temporal order, and distinctive shared features.

A child concept can have a parent set rather than one parent. The transformation is therefore represented by a directed hyperedge:

{C(p1), C(p2), …, C(pm)} -> C(j)

Branching occurs when one earlier version contributes to two later versions that become increasingly dissimilar. Fusion occurs when a later version inherits substantial structure from more than one lineage. Differentiation can be quantified using a combination of retained similarity to the ancestor, declining similarity between descendants, increasing context selectivity, and divergence in predictive consequences.

Artificial Intelligence and Cognitive Provenance

An artificial system based on iterative updating could record its conceptual ancestry directly. Each addition to its workspace could retain weighted provenance links to the active representations, retrieved memories, external sources, and prediction errors that contributed to its selection. When a sequence produces a durable memory update, the system could store a versioned record of the affected representation and its parent set.

This would create a cognitive provenance ledger. Unlike a retrospective reconstruction from human writings, the ledger could capture internal transformations as they occur. It could distinguish a newly inferred relationship from a retrieved one, identify which earlier representations were jointly necessary for a synthesis, and show when a concept divided into context-specific descendants.

Causal intervention would improve the record. If removing an alleged ancestor leaves a descendant unchanged, the connection may reflect superficial similarity. If ablation reliably changes the descendant’s probability or structure, the ancestry claim gains causal support. The system could also compare checkpoints of its memory network and attribute changes to particular iterative episodes.

Exact provenance would aid explainability and scientific discovery. A system could present the developmental pathway of a conclusion, revisit a branch point, restore a discarded alternative, or identify the external source from which a premise entered its memory. It could also detect when two apparently independent conclusions share a hidden ancestor or when similar solutions emerged through convergence.

The design would introduce costs. Recording every internal transition would produce enormous histories, and some provenance information could expose private data or copyrighted sources. Practical systems would need salience thresholds, compressed checkpoints, source permissions, and selective retention. The theoretical principle remains straightforward: an intelligence that changes its own long-term memory should be able to preserve evidence about how consequential changes occurred.

Boundaries of the Evolutionary Analogy

Conceptual phylogeny should not be mistaken for a claim that ideas are biological organisms. Neural representations do not possess genomes, reproduce as autonomous individuals, or compete in a single well-defined population. Acquired changes can be preserved directly, goals can guide variation, and one descendant routinely inherits from many parents. These features make the process partly Lamarckian, highly reticulate, and strongly dependent on agency.

The boundaries do not eliminate the underlying historical structure. Evolution in its broadest sense requires variation, persistence, differential retention, and inherited modification. Conceptual development exhibits these properties when earlier representations causally constrain later ones. Phylogenetic language is valuable to the extent that it sharpens measurable distinctions among shared ancestry, convergence, branching, and recombination.

Several methodological problems remain. The boundaries of an idea are observer-dependent. Documents sample only expressed cognition. Current neural measures are coarse relative to distributed assemblies. Autobiographical reports can be reconstructed in light of current beliefs. Similarity measures can mistake common vocabulary for descent. Any serious reconstruction must preserve uncertainty and allow several candidate histories when the evidence does not decide among them.

There is also a risk of assuming the process that the reconstruction is intended to demonstrate. A branching diagram drawn from semantically clustered texts does not prove that corresponding neural representations branched. Behavioral, documentary, computational, and neural evidence should be compared rather than collapsed. The strongest tests will prospectively define candidate ancestral structures and observe how they change under controlled learning conditions.

Conclusion

The iterative updating model explains how one mental state develops from another through partial retention, associative search, and incremental replacement. Its learning component implies a longer history. Coactive states modify the network that will generate subsequent states. Later retrieval reinstates selected products, exposes them to new contexts, and sometimes preserves a modified descendant. Repetition of this cycle produces lineages of conceptual change.

These lineages differentiate, merge, compress, disappear, and return. Their organization is usually more accurately represented by a phylogenetic network than by a tree. Ancestry resides in causal continuity and inherited relational structure rather than exact repetition. External records provide partial fossils of the process and can also reactivate old branches. Timestamped intellectual corpora therefore offer an empirical route for reconstructing conceptual development, while longitudinal neuroscience can test the predicted changes in representational geometry and retrieval dynamics.

This extension connects momentary thought to lifelong intellectual development. Working memory supplies the local arena in which variants are generated and combined. Learning preserves selected changes in long-term memory. Recurrent retrieval turns those preserved changes into historical lineages. A phylogeny of ideas is the accumulated record of iterative thought acting upon the memory system that makes further thought possible.

References

Anderson, J. R. (1983). A spreading activation theory of memory. Journal of Verbal Learning and Verbal Behavior, 22(3), 261-295. https://doi.org/10.1016/S0022-5371(83)90201-3

Bartlett, F. C. (1932). Remembering: A study in experimental and social psychology. Cambridge University Press.

Campbell, D. T. (1960). Blind variation and selective retention in creative thought as in other knowledge processes. Psychological Review, 67(6), 380-400. https://doi.org/10.1037/h0040373

Collins, A. M., & Loftus, E. F. (1975). A spreading activation theory of semantic processing. Psychological Review, 82(6), 407-428. https://doi.org/10.1037/0033-295X.82.6.407

Cowan, N. (1988). Evolving conceptions of memory storage, selective attention, and their mutual constraints within the human information processing system. Psychological Bulletin, 104(2), 163-191. https://doi.org/10.1037/0033-2909.104.2.163

Damasio, A. R. (1989). Time-locked multiregional retroactivation: A systems-level proposal for the neural substrates of recall and recognition. Cognition, 33(1-2), 25-62. https://doi.org/10.1016/0010-0277(89)90005-X

Fuster, J. M. (2009). Cortex and memory: Emergence of a new paradigm. Journal of Cognitive Neuroscience, 21(11), 2047-2072. https://doi.org/10.1162/jocn.2009.21280

Gray, R. D., & Atkinson, Q. D. (2003). Language tree divergence times support the Anatolian theory of Indo European origin. Nature, 426, 435-439. https://doi.org/10.1038/nature02029

Greenhill, S. J., Currie, T. E., & Gray, R. D. (2009). Does horizontal transmission invalidate cultural phylogenies? Proceedings of the Royal Society B, 276(1665), 2299-2306. https://doi.org/10.1098/rspb.2008.1944

Hebb, D. O. (1949). The organization of behavior. Wiley.

Howard, M. W., & Kahana, M. J. (2002). A distributed representation of temporal context. Journal of Mathematical Psychology, 46(3), 269-299. https://doi.org/10.1006/jmps.2001.1388

Hupbach, A., Gomez, R., Hardt, O., & Nadel, L. (2007). Reconsolidation of episodic memories: A subtle reminder triggers integration of new information. Learning and Memory, 14(1-2), 47-53. https://doi.org/10.1101/lm.365707

McClelland, J. L., McNaughton, B. L., & O’Reilly, R. C. (1995). Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory. Psychological Review, 102(3), 419-457. https://doi.org/10.1037/0033-295X.102.3.419

Miller, E. K., & Cohen, J. D. (2001). An integrative theory of prefrontal cortex function. Annual Review of Neuroscience, 24, 167-202. https://doi.org/10.1146/annurev.neuro.24.1.167

Mongillo, G., Barak, O., & Tsodyks, M. (2008). Synaptic theory of working memory. Science, 319(5869), 1543-1546. https://doi.org/10.1126/science.1150769

Nader, K., Schafe, G. E., & LeDoux, J. E. (2000). Fear memories require protein synthesis in the amygdala for reconsolidation after retrieval. Nature, 406, 722-726. https://doi.org/10.1038/35021052

Reser, J. E. (2016). Incremental change in the set of coactive cortical assemblies enables mental continuity. Physiology and Behavior, 167, 222-237. https://doi.org/10.1016/j.physbeh.2016.09.019

Reser, J. E. (2024). A cognitive architecture for machine consciousness and artificial superintelligence: Thought is structured by the iterative updating of working memory. arXiv:2203.17255v7. https://doi.org/10.48550/arXiv.2203.17255

Rose, N. S., LaRocque, J. J., Riggall, A. C., Gosseries, O., Starrett, M. J., Meyering, E. E., & Postle, B. R. (2016). Reactivation of latent working memories with transcranial magnetic stimulation. Science, 354(6316), 1136-1139. https://doi.org/10.1126/science.aah7011

Schlichting, M. L., Mumford, J. A., & Preston, A. R. (2015). Learning related representational changes reveal dissociable integration and separation signatures in the hippocampus and prefrontal cortex. Nature Communications, 6, 8151. https://doi.org/10.1038/ncomms9151

Schlichting, M. L., & Preston, A. R. (2015). Memory integration: Neural mechanisms and implications for behavior. Current Opinion in Behavioral Sciences, 1, 1-8. https://doi.org/10.1016/j.cobeha.2014.07.005

Tehrani, J. J. (2013). The phylogeny of Little Red Riding Hood. PLOS ONE, 8(11), e78871. https://doi.org/10.1371/journal.pone.0078871

Triana, A. M., Salmi, J., Hayward, N. M. E. A., Saramaki, J., & Glerean, E. (2024). Longitudinal single subject neuroimaging study reveals effects of daily environmental, physiological, and lifestyle factors on functional brain connectivity. PLOS Biology, 22(10), e3002797. https://doi.org/10.1371/journal.pbio.3002797

Zeithamova, D., Dominick, A. L., & Preston, A. R. (2012). Hippocampal and ventral medial prefrontal activation during retrieval mediated learning supports novel inference. Neuron, 75(1), 168-179. https://doi.org/10.1016/j.neuron.2012.05.025