Semantic-Group Shared Interest Memory

• Semantic-Group Shared Interest Memory is a framework that explains how groups collectively externalize, stabilize, and update semantic structures via language.
• It leverages decentralized Bayesian inference in a Collective Predictive Coding system where individual and group-level processes reduce prediction error.
• The framework integrates internal memory with external language-based representations to enable dynamic updating and coordinated attention to shared interests.

Searching arXiv for the primary paper and closely related memory papers to ground the article. Semantic-Group Shared Interest Memory is a conceptual account of how shared semantic structures, interests, and knowledge can be formed, maintained, and updated at the level of a collective rather than an individual. In the framework of Collective Predictive Coding (CPC), language is treated as a collectively formed external representation through which a group of agents can be viewed as a single higher-level predictive-coding system, with memory and attention generalized to the collective scale (Taniguchi, 20 Aug 2025). In this sense, Semantic-Group Shared Interest Memory denotes the evolving, collectively maintained semantic space that reflects what a group has learned, cares about, and expects, with language and related symbolic artifacts functioning as the persistent substrate of this group-level memory (Taniguchi, 20 Aug 2025).

1. Collective Predictive Coding and group-level memory

The core formulation of CPC distinguishes two coupled levels of inference. At the individual level, each agent performs internal representation learning or world modeling. At the collective level, the group performs external representation learning, identified with language or symbol emergence. Both levels are treated as decentralized Bayesian inference: individual agents minimize their own free energy under their own generative models, while the collective performs approximate Bayesian inference over shared symbols and linguistic structures through communicative interactions such as naming games (Taniguchi, 20 Aug 2025).

A central claim is that “an agent collective, coupled by the emergent language, can be viewed as a single entity performing active inference or predictive coding” (Taniguchi, 20 Aug 2025). CPC therefore extends standard predictive coding by positing a two-layer inferential organization: an internal layer in which each agent maintains its own generative model $p_i(o_i, s_i)$, and an external or collective layer in which the population collectively infers and updates a shared symbolic system. This shared symbolic system functions as an external generative model over communicative symbols and, implicitly, over the world (Taniguchi, 20 Aug 2025).

The commentary further invokes the “System 3” notion, extending Kahneman’s System 1 and System 2 so that society, or a scientific community, becomes the subject of active inference. In that picture, the collective has aggregated observations, an effective generative model encoded in language, texts, theories, and norms, and coordinated forms of action that tend to reduce group-level prediction error (Taniguchi, 20 Aug 2025). The paper notes that Taniguchi et al. (2025) derive the free energy of a human collective as a model of science and describe a collective regularization term representing “semiotic plasticity,” meaning that the collective generative model in language and scientific theories is molded to minimize mismatch with individuals’ experiences while individuals are simultaneously constrained by shared semiotic structures (Taniguchi, 20 Aug 2025).

This suggests that Semantic-Group Shared Interest Memory is not merely a metaphor for collective knowledge. Within CPC it is the external representational layer by which a group stabilizes, regularizes, and updates a shared model of $p(\mathbf{o})$, the distribution of the group’s observations (Taniguchi, 20 Aug 2025).

2. Memory and attention beyond the individual

The commentary explicitly characterizes language as enabling external, collective memory: “through written language, society can store information externally, forming a collective memory. Individuals can also access memories stored by others through communication” (Taniguchi, 20 Aug 2025). In this account, memory is not limited to neural traces. It includes written texts, scientific publications, laws, digital archives, linguistic conventions, and commonly used terms that stabilize concepts across time (Taniguchi, 20 Aug 2025).

Within CPC, individual-level memory corresponds to an agent’s internal generative model, including learned parameters and latent structure, whereas collective-level memory corresponds to the public representational space formed by language and symbolic structures (Taniguchi, 20 Aug 2025). That public representational space stores semantic regularities across time and across individuals, allowing information to be offloaded to external media and later re-accessed by other members of the collective (Taniguchi, 20 Aug 2025). Semantic-Group Shared Interest Memory is therefore the collective-level counterpart of internal predictive memory.

Collective attention is treated as the factorization of group-level cognition into individual attention mechanisms. The commentary suggests that the collective, considered as System 3, is factorized into individual beings, each with their own attentional mechanisms; collective attention is then the distribution of attentional resources across agents and topics, shaped by the collective generative model, including language, norms, and shared goals (Taniguchi, 20 Aug 2025). In predictive-coding terms, this is naturally interpreted as modulation of precision across communicative channels and topics: topics receiving frequent discussion, more data, and more refinement have high effective precision, whereas neglected topics fade from collective memory and receive fewer updates (Taniguchi, 20 Aug 2025).

This supports the idea that shared interests are not separable from memory. The same semantic structures that encode what a group knows also encode where the group allocates precision and update capacity. A plausible implication is that Semantic-Group Shared Interest Memory is simultaneously a memory substrate and an attentional policy over that substrate.

3. Language as external representation and collective world model

The central thesis of the CPC commentary is that “language, with its embedded distributional semantics, serves as a collectively formed external representation” (Taniguchi, 20 Aug 2025). Internal representation learning occurs within individuals; external representation learning occurs at the population level through communication, yielding a symbol system that encodes regularities across shared observations (Taniguchi, 20 Aug 2025).

The mechanism cited for symbol emergence is decentralized Bayesian inference approximated by language games, including the Metropolis–Hastings Naming Game, where agents propose labels for observations and these proposals are accepted or rejected according to a probabilistic rule akin to Metropolis–Hastings sampling (Taniguchi, 20 Aug 2025). Over time, a shared lexicon emerges that best explains the joint data. This makes language emergence a process of decentralized inference over the space of possible symbol systems, driven by the reduction of collective prediction error (Taniguchi, 20 Aug 2025).

The commentary places special emphasis on distributional semantics and embedding spaces. It states that embedding spaces are rooted in linguistic structures that give rise to distributional semantics, and asks who created the linguistic structures manifesting in those spaces; its answer is that these structures are created not only by individual humans but by humans as a collective (Taniguchi, 20 Aug 2025). In CPC terms, word co-occurrence statistics and embedding spaces compress patterns of word use produced by many individuals over time. Those distributed representations therefore encode the group’s shared semantic associations and world knowledge (Taniguchi, 20 Aug 2025).

The paper also answers the question “Is next-word prediction the most effective way to learn language?” with a CPC-based affirmative argument. It states that next-token prediction is fundamentally about modeling the probability distribution of observations, $p(\mathbf{o})$, and if language integrates the observations of a group of partially observable agents and structurally represents $p(\mathbf{o})$, then modeling this distribution is the essence of human language learning and the learning of a collective world model (Taniguchi, 20 Aug 2025). On this interpretation, LLMs approximate the collective world model encoded in language. Their token embeddings and hidden states can thus be understood as approximations of the group’s shared world model and shared interest structure (Taniguchi, 20 Aug 2025).

This suggests that Semantic-Group Shared Interest Memory can be operationalized as a statistical structure over linguistic outputs. Distributional semantics and next-token models do not merely compress language use; they instantiate an externalized memory of collective expectations, associations, and interests.

4. Formation, maintenance, and updating of shared semantic interests

Although the phrase “Semantic-Group Shared Interest Memory” does not appear verbatim in the CPC commentary, the paper’s description aligns closely with it (Taniguchi, 20 Aug 2025). Semantic structures, categories, roles, relationships, and recurrent topics are said to arise as stable patterns in language and to be collectively maintained through repeated use (Taniguchi, 20 Aug 2025). Shared interests and goals are reflected in recurring topics, specialized terminology, and normative vocabularies that encode group priorities and constraints (Taniguchi, 20 Aug 2025).

The commentary describes a three-part dynamic. Formation occurs through decentralized Bayesian inference, with individuals proposing new words, senses, or constructions that are accepted, rejected, or modified according to communicative success or failure (Taniguchi, 20 Aug 2025). Maintenance occurs because successful communicative patterns reduce prediction error and are reinforced through repeated interactions and through their inscription into external media such as texts, code, and databases (Taniguchi, 20 Aug 2025). Updating occurs when mismatches arise between existing linguistic structures and new experiences; prediction error at both individual and collective levels then drives the introduction of new terms, shifts in usage, and alterations in narrative structures (Taniguchi, 20 Aug 2025).

The notion of “semiotic plasticity,” mentioned in relation to collective free energy, captures the collective capacity to change its symbol system so that words, meanings, and linguistic structures better fit the evolving world and the changing distribution of group observations and interests (Taniguchi, 20 Aug 2025). Semantic-Group Shared Interest Memory is therefore dynamic rather than archival. It is continually restructured by predictive-coding processes operating across both individual and collective scales (Taniguchi, 20 Aug 2025).

The commentary also indicates that heterogeneity is intrinsic to the process. Agents have different internal models and partial observations; decentralized inference implies that sub-groups may specialize in different parts of the generative model, with local terminologies and perspectives, while higher-level language supports coordination across sub-groups (Taniguchi, 20 Aug 2025). This implies a layered structure of shared memory, with globally shared semantics and locally specialized semantics coexisting within a single collective representational system (Taniguchi, 20 Aug 2025).

Norms and conventions occupy a central position in this account. Scientific norms about evidence, and moral or legal norms about behavior and consequences, are treated as priors in the collective predictive-coding system (Taniguchi, 20 Aug 2025). They guide individual inferences and shape the space of likely utterances, thereby encoding shared interests in communication and coordination (Taniguchi, 20 Aug 2025).

5. Related memory formalisms and computational analogues

Several other memory frameworks illuminate different computational aspects of Semantic-Group Shared Interest Memory. Sparsey provides a model in which episodic memory and semantic memory are not separated into distinct modules but are realized in the same superposed sparse distributed representations (SDRs) (Rinkus et al., 2017). In Sparsey, similar inputs map to codes with high intersection, while dissimilar inputs map to codes with low intersection, so the geometry of overlap in code space mirrors the geometry of similarity in input space (Rinkus et al., 2017). This means that semantic group structure emerges from patterns of overlap among codes rather than from explicit prototypes (Rinkus et al., 2017).

In the language of Semantic-Group Shared Interest Memory, Sparsey treats semantic groups as high-overlap neighborhoods in SDR space. A user profile or an item can be represented as a set of SDRs across hierarchical coding fields, and shared interest between two users can be computed as the intersection of unions of those episode codes (Rinkus et al., 2017). This suggests a memory architecture in which shared interests are encoded directly in overlapping distributed codes, with no separate semantic store. The paper argues that the same superposed structure supports both episodic retrieval and semantic generalization, making semantic memory a “computationally free side-effect” of storing episodic traces (Rinkus et al., 2017).

A different computational analogue appears in semantic communication with shared prototype memories. In that framework, transmitter and receiver maintain synchronized semantic memories $$\mathcal{M}_t = \{\boldsymbol{\mu}_1,\ldots,\boldsymbol{\mu}_{M_t}\}$$, where each prototype represents a semantic concept such as “blue sky,” “static background,” or “person walking” (Nasreddine et al., 13 Nov 2025). Modern Hopfield networks perform soft attention-based retrieval over those prototypes, allowing smooth reuse of stored concepts under bounded semantic drift (Nasreddine et al., 13 Nov 2025). The paper explicitly interprets each prototype as a semantic group center, with memory reuse rate and compression ratio combined into a semantic efficiency measure and a “reasoning capacity” metric (Nasreddine et al., 13 Nov 2025).

This communication setting differs from CPC in scope, but it supports the same general intuition: a shared semantic memory can be organized around recurring semantic groups, and soft retrieval can preserve stable group assignments even as observations evolve (Nasreddine et al., 13 Nov 2025). A plausible implication is that Semantic-Group Shared Interest Memory in collective cognition may likewise depend on graded rather than purely hard assignment to semantic prototypes.

6. Implementations in recommendation, agent memory, and anticipatory recall

Recent systems in recommendation and agent memory make the notion of group-shared semantic memory more explicit. ISRF, designed for generative recommendation with LLMs, distinguishes explicit individual interests from implicit group interests (Zhu et al., 14 Mar 2026). It performs multi-step bidirectional reasoning over item attributes to infer semantic item features, constructs a semantic interaction graph to model explicit interests, and builds a similarity-based user graph to infer implicit interests of similar user groups (Zhu et al., 14 Mar 2026). Group-level interest representations are obtained by applying LightGCN to the user graph, producing embeddings $\mathbf{H}$ in which each row $\mathbf{h}_u$ is a group-level interest representation for user $u$ (Zhu et al., 14 Mar 2026).

The paper explicitly states that $\mathbf{S}_u, \mathcal{R}, \mathbf{H}$ collectively form an external memory bank of semantic-group preferences and that $\mathbf{H}$ can be interpreted as a latent, distributed memory of shared interest patterns across the user graph (Zhu et al., 14 Mar 2026). Through contrastive objectives that align individual and group representations, explicit interests guide refinement of group implicit interests and group implicit interests enhance individual modeling (Zhu et al., 14 Mar 2026). This offers a concrete graph-based realization of Semantic-Group Shared Interest Memory in recommendation systems.

CLAG provides an agent-memory architecture organized as semantic neighborhoods rather than a single global retrieval pool (Roh et al., 16 Mar 2026). Memory notes are assigned to clusters by an SLM-driven router, and each cluster has a centroid and a profile consisting of a one-sentence summary and three descriptive tags (Roh et al., 16 Mar 2026). Each cluster is treated as a self-contained functional unit, and both localized evolution and retrieval occur primarily within clusters (Roh et al., 16 Mar 2026). The framework thus implements semantic-group memory by making each cluster a semantically coherent shared-interest unit.

T-Mem addresses a different problem: long-term conversational memory that must support both descriptive recall and associative recall (Guo et al., 13 Jun 2026). It introduces write-time trigger generation so that every memory remains reachable not only from surface-similar queries but also from future queries connected by latent semantic arcs (Guo et al., 13 Jun 2026). Its memory structure combines topics, scenes, items, and four trigger families: Entity, Bridge, Scene, and Horizon triggers (Guo et al., 13 Jun 2026). In the terminology of Semantic-Group Shared Interest Memory, topics provide explicit groupings of recurring semantic threads, while associative triggers function as semantic-group access paths that connect later queries to earlier memories even when lexical overlap is absent (Guo et al., 13 Jun 2026).

These systems differ in ontology and mechanism, but all of them move beyond unstructured storage. They organize shared memory by semantic group structure, whether through distributed overlap patterns, prototype memories, user graphs, semantic clusters, or anticipatory triggers.

7. Shared external memory without explicit communication

A distinct perspective comes from work on environmental memory in active matter, where group formation emerges among “clueless individuals” that have no explicit communication and no internal memory (Dias et al., 2023). In that system, active Janus colloids move through a dynamic environment of passive colloidal obstacles. As active particles traverse the environment, they mechanically displace obstacles and open transient low-density paths. Those paths persist for a finite time and bias the motion of later particles, which preferentially reuse them (Dias et al., 2023).

The environment thus stores a transient external memory of where particles have recently passed. The paper interprets this as a mechanism of indirect coordination via shared memory that can overshadow the need for explicit signaling (Dias et al., 2023). Although the domain is physical rather than linguistic, the conceptual relevance is direct: collective coordination can be mediated by a shared external substrate that records traces of prior activity and biases future action. A plausible implication is that Semantic-Group Shared Interest Memory need not be linguistic in every implementation. Any dynamic external substrate that stores semantically relevant traces and feeds them back into group behavior can serve an analogous role.

In collective cognition, language, texts, archives, embeddings, graphs, and symbolic media play the role that obstacle configurations play in the active-matter system. They are shared, persistent, partially decaying structures that encode historical activity and channel future coordination. The CPC commentary’s treatment of language as a collectively formed external representation is therefore consistent with a broader class of shared-memory mechanisms in which the environment itself becomes the memory substrate (Taniguchi, 20 Aug 2025).

8. Significance, scope, and open interpretation

Semantic-Group Shared Interest Memory provides a framework for understanding how shared semantics, shared interests, and group-level cognition can be stabilized in publicly accessible representational media. In CPC, this framework is centered on language as an external generative model that integrates observations from many partially observable agents and represents $p(\mathbf{o})$0 at the collective scale (Taniguchi, 20 Aug 2025). In related computational systems, the same underlying idea appears as overlap structure in SDRs, synchronized semantic prototypes, user-graph embeddings, agentic semantic clusters, or trigger-based long-term memory (Rinkus et al., 2017, Nasreddine et al., 13 Nov 2025, Zhu et al., 14 Mar 2026, Roh et al., 16 Mar 2026, Guo et al., 13 Jun 2026).

A common misconception is that shared semantic memory is merely a repository of stored content. The cited work points instead to a coupled memory-attention system in which what is stored, what is reachable, and what is updated all depend on recurring communicative use, precision allocation, and mechanisms of semantic reuse (Taniguchi, 20 Aug 2025). Another misconception is that group-level cognition requires explicit centralized control. The CPC commentary and the active-matter environmental-memory study both support the opposite possibility: group-level structure can emerge through decentralized interaction with shared external representations (Taniguchi, 20 Aug 2025, Dias et al., 2023).

The strongest claims presently available are conceptual rather than universally formalized. The CPC commentary is explicit that much of the machinery is referenced rather than fully derived (Taniguchi, 20 Aug 2025). Accordingly, the most rigorous statement is that Semantic-Group Shared Interest Memory names a family of mechanisms by which collectives externalize, stabilize, and update shared semantic structures and shared priorities. Within that family, language occupies a privileged role because it is both memory and model: a persistent symbolic substrate through which a group can learn, communicate, and act as if it were a higher-level predictive-coding system (Taniguchi, 20 Aug 2025).