Subjective Experience Memory Mechanisms

• Subjective Experience-Based Memory Mechanism is defined as a memory system organized around salient, self-relevant, and affectively weighted events that guide recollection and planning.
• It encompasses diverse processes such as episodic re-experiencing, salience-driven temporal segmentation, and appraisal-based reinforcement that have been modeled both neurally and computationally.
• The framework informs both neuroscience and artificial-agent architectures, enhancing our understanding of memory’s role in decision-making and performance optimization.

Searching arXiv for the cited papers to ground the article. arxiv_search: query="(Zakharov et al., 2020) Episodic Memory for Learning Subjective-Timescale Models", max_results=5 Subjective experience-based memory mechanisms encompass memory processes in which storage, organization, or retrieval is shaped by the properties of lived experience rather than by undifferentiated objective time or flat logging. In the literature, this includes episodic memory as “the ability to re-experience a time and place specific event in its original context,” subjective-timescale dynamics learned from salient episodic memories, appraisal states updated by competing positive and negative reinforcement memories, and agent memories built from actions, observations, feedback, and reflection (Pastor, 2020, Zakharov et al., 2020, Hu et al., 2022, Zhang et al., 2024). The mechanisms differ in substrate and formalism, but they converge on a common operational theme: memory is organized around what is salient, self-relevant, affectively weighted, temporally structured, or behaviorally useful, and is then reused to guide recollection, planning, and decision-making.

1. Conceptual scope and taxonomy

Within the multiple-memory-systems framework, episodic memory is the paradigmatic subjective memory system. It is defined as “the ability to re-experience a time and place specific event in its original context,” and its hallmark is mental time travel through “time, space and sense of self” (Pastor, 2020). This stands in contrast to semantic memory, described as “a mental thesaurus encompassing a wide range of organized information including facts, concepts and vocabulary,” and to nondeclarative memory, which is “expressed through performance rather than recollection” and is typically unconscious (Pastor, 2020).

Artificial-agent work uses a parallel distinction. In LLM-based agents, episodic memory corresponds to concrete action–observation histories, trial-level outcomes, and environmental feedback, whereas semantic memory corresponds to consolidated summaries, schemas, rules, and traits distilled from those episodes (Zhang et al., 2024). The basic operational pipeline is formalized as

$m_{t}^k = W(\{a^k_t, o^k_t\}),$

$M_{t}^k = P(M_{t-1}^k, m_{t}^k),$

$\hat{M}_{t}^k = R(M_{t}^k, c_{t+1}^k),$

with the next action conditioned on retrieved memory (Zhang et al., 2024).

A broader taxonomy follows from these definitions. Some mechanisms emphasize recollective subjectivity and autobiographical context; others emphasize salience-defined temporal segmentation, appraisal polarity, or distilled procedural know-how. This suggests that “subjective experience-based memory mechanism” is best treated as an umbrella category spanning at least four recurring motifs: episodic re-experiencing, salience-gated event memory, affective or appraisal-based reinforcement memory, and agent-centric experience memory.

2. Neural and cognitive foundations of subjective memory

Episodic memory depends on the hippocampus and adjoining medial temporal lobe structures, including entorhinal, perirhinal, and parahippocampal cortices, which together support item and context representation and associative binding (Pastor, 2020). The paper specifies a functional differentiation inside the hippocampal circuit: dentate gyrus supports sparse orthogonalization and pattern separation, CA3 supports auto-associative pattern completion, and CA1 compares and integrates CA3 outputs with direct entorhinal inputs, supporting temporal-order and mismatch detection (Pastor, 2020).

Retrieval is not treated as a simple replay of stored content. Rather, hippocampal indexing and cortical reinstatement reconstruct an episode from distributed traces. Computationally, this is expressed by an indexing scheme such as

$h = W_e x,$

at encoding, and

$h = W_e \tilde{x}, \qquad \hat{x} = W_c h,$

at retrieval, with Hebbian binding

$\Delta w_{ij} = \eta x_i x_j$

supporting item–context associations (Pastor, 2020). CA3-like pattern completion is likewise formalized with attractor dynamics, making recollection a constructive reassembly rather than a literal playback (Pastor, 2020).

Prefrontal and parietal systems modulate this process. The prefrontal cortex supports strategic encoding, retrieval, and source memory, and frontal lesions can produce source amnesia, in which remembered facts are stripped of spatiotemporal context (Pastor, 2020). The same literature also emphasizes overlap between remembering and imagining future episodes, including common hippocampal involvement, which links subjective memory to prospective simulation rather than to a purely retrospective store (Pastor, 2020). This is consistent with the claim that autonoetic consciousness is not only about access to stored content, but about reconstructing a self in subjective time.

3. Salience, event boundaries, and subjective temporality

One line of work replaces objective clock time with a salience-defined subjective timescale. In the subjective-timescale model, the agent accumulates episodic memories into sequences

$$S_e = \{s_{\tau_1}, s_{\tau_2}, \ldots, s_{\tau_N}\},$$

where subjective indices $\tau$ are determined not by every objective step but by salient transitions (Zakharov et al., 2020). Saliency is computed from transition-model prediction error,

$$S_t = D_{KL}[ q(s_t; \phi_s) \,\|\, p(s_t \mid s_{t-1}, a_{t-1}; \theta_s) ],$$

and a memory is formed when $S_t > \epsilon$ (Zakharov et al., 2020). The resulting transition model predicts between event boundaries,

$M_{t}^k = P(M_{t-1}^k, m_{t}^k),$0

instead of rolling forward at unit time steps (Zakharov et al., 2020).

This mechanism is motivated by the claim that one-step objective roll-outs accumulate error and waste computation on slow dynamics. By predicting “keyframes” at event boundaries, the model reduces one-step error accumulation, skips irrelevant intermediate states, and adapts temporal granularity without explicit horizon control (Zakharov et al., 2020). In Animal-AI, the STM-MCTS agent achieved higher cumulative reward than Baseline-MCTS and surpassed Baseline-MPC despite MPC’s higher computation, reaching greater reward in approximately 6 hours versus approximately 14 hours for MPC (Zakharov et al., 2020).

A related but biologically framed account is the synaptic clock theory, which proposes that every synapse carries its own characteristic time unit set by the persistence of short-lasting memory traces. The minimal trace dynamics are written as

$M_{t}^k = P(M_{t-1}^k, m_{t}^k),$1

and perceived duration scales with the number of subjectively detected events,

$M_{t}^k = P(M_{t-1}^k, m_{t}^k),$2

(Jura, 2018). Shorter effective time constants increase temporal resolution and can yield time dilation; longer ones promote event fusion and time compression (Jura, 2018).

A third formulation appears in the self-simulational theory of temporal extension, where the width of the subjective temporal Now is identified with the risk term of expected free energy,

$M_{t}^k = P(M_{t-1}^k, m_{t}^k),$3

Its experiential memory mechanism is a precision-weighted retention/protention buffer implemented by smoothing,

$M_{t}^k = P(M_{t-1}^k, m_{t}^k),$4

so that the just-past and near-future jointly determine present subjective temporality (Bellingrath, 2024). This suggests that subjective memory and subjective time are tightly coupled wherever event segmentation, prediction error, or counterfactual policy evaluation determines what counts as a meaningful temporal unit.

4. Appraisal, density tagging, and prioritization signals

Another family of mechanisms treats memory as explicitly shaped by affective polarity or appraisal competition. The Reinforcement Memory Unit maintains two reinforcement memories, $M_{t}^k = P(M_{t-1}^k, m_{t}^k),$5 and $M_{t}^k = P(M_{t-1}^k, m_{t}^k),$6, together with an appraisal state $M_{t}^k = P(M_{t-1}^k, m_{t}^k),$7 (Hu et al., 2022). A stimulus response is computed as

$M_{t}^k = P(M_{t-1}^k, m_{t}^k),$8

forget gates attenuate prior memory,

$M_{t}^k = P(M_{t-1}^k, m_{t}^k),$9

and the memories update through stronger-stimulus replacement,

$\hat{M}_{t}^k = R(M_{t}^k, c_{t+1}^k),$0

$\hat{M}_{t}^k = R(M_{t}^k, c_{t+1}^k),$1

followed by competitive appraisal,

$\hat{M}_{t}^k = R(M_{t}^k, c_{t+1}^k),$2

(Hu et al., 2022). The paper reports that RMU generally surpasses LSTM and GRU, particularly for BRISQUE, VBLINDS, and TLVQM on correlation metrics, and attains the lowest RMSE and highest PLCC on the QoE task (Hu et al., 2022).

A distinct prioritization architecture appears in Modular Consciousness Theory. There, consciousness is a discrete sequence of Integrated Informational States, each annotated with a multidimensional information-density vector; memory “uses the vector to prioritize storage and consolidation,” and higher vector amplitude increases the probability that the IIS will be stored and later retrieved (Gillon, 2 Oct 2025). The theory does not provide an explicit function from density magnitude to write strength, but it states a monotonic relation between amplitude and memory priority, as well as between amplitude and behavioral or decision influence (Gillon, 2 Oct 2025). The same theory therefore reframes subjectivity as an internal tagging signal with functional consequences rather than as an unanalyzed phenomenal residue.

Stimulus-driven memorability introduces a related, but not identical, perspective. Certain items are intrinsically memorable across observers, and memory performance can be approximated by a combined model

$\hat{M}_{t}^k = R(M_{t}^k, c_{t+1}^k),$3

where $\hat{M}_{t}^k = R(M_{t}^k, c_{t+1}^k),$4 is intrinsic memorability, $\hat{M}_{t}^k = R(M_{t}^k, c_{t+1}^k),$5 aggregates subjective-state variables, and $\hat{M}_{t}^k = R(M_{t}^k, c_{t+1}^k),$6 captures task or context (Bainbridge, 2021). The chapter states that the image factor explains more than half the variance in memory performance, and that reward and attention manipulations do not reverse intrinsic item rankings (Bainbridge, 2021). This yields an important qualification: subjective experience modulates memory, but in some domains it operates atop a strong stimulus-driven backbone rather than replacing it.

5. Artificial-agent implementations and memory-centric architectures

In agent systems, subjective experience is often operationalized as the agent’s own interaction history together with later reflection or abstraction. The LLM-agent survey formalizes memory as a write–process–read pipeline over inside-trial information, cross-trial experiences, and external knowledge contextualized through the agent’s own state (Zhang et al., 2024). The later evolutionary survey sharpens this into three stages—Storage, Reflection, and Experience—where trajectories are first preserved, then refined,

$\hat{M}_{t}^k = R(M_{t}^k, c_{t+1}^k),$7

and finally abstracted across trajectories into a compact prior,

$\hat{M}_{t}^k = R(M_{t}^k, c_{t+1}^k),$8

with the explicit requirement that $\hat{M}_{t}^k = R(M_{t}^k, c_{t+1}^k),$9 (Luo et al., 7 May 2026). This stage-wise account treats subjective memory not as passive accumulation but as a lifecycle of preservation, distillation, and reuse.

Concrete architectures instantiate this principle in different ways. Memoir uses a language-conditioned world model to imagine future navigation states and uses those imagined states as retrieval queries for a Hybrid Viewpoint-Level Memory containing both observation histories and navigation behavioral histories (Xu et al., 9 Oct 2025). On IR2R, Memoir improves Val-Unseen SPL from 67.9% to 73.3%, corresponding to a 5.4% SPL gain over the best memory-persistent baseline, while also reporting 8.3× training speedup and 74% inference memory reduction (Xu et al., 9 Oct 2025). The key design move is that the agent retrieves not only what was seen but also how it previously reasoned and moved, so that memory becomes explicitly agent-centric.

Other systems emphasize explicit structure and provenance. DGMM defines memory as a typed graph

$h = W_e x,$0

updates it additively,

$h = W_e x,$1

and constructs working memory by cue-conditioned recall,

$h = W_e x,$2

(Dorsey et al., 4 May 2026). Episodic persistence and contextual variability are central invariants: stored memory is not modified by recall-conditioned interpretation, but different cues can yield different working subgraphs and different downstream propositions (Dorsey et al., 4 May 2026). ReMe, by contrast, stores procedural experiences as

$h = W_e x,$3

and prunes them with a utility rule based on successful contributions,

$h = W_e x,$4

with $h = W_e x,$5 and $h = W_e x,$6 in experiments (Cao et al., 11 Dec 2025).

Relational and residual designs push the same idea further. GSEM organizes clinical experiences into a dual-layer graph with experience qualities $h = W_e x,$7 and inter-experience weights $h = W_e x,$8, then updates both online from outcome feedback (Han et al., 23 Mar 2026). DeltaMem organizes experience into two residual trees, one for skills and one for environment knowledge, and reconstructs full experience by root-to-match composition,

$h = W_e x,$9

for skills and

$h = W_e \tilde{x}, \qquad \hat{x} = W_c h,$0

for knowledge (Tan et al., 2 Jun 2026). This suggests that recent agent architectures increasingly treat subjective experience not as a flat archive, but as a structured, queryable, and self-revising substrate.

6. Debates, misconceptions, and open problems

A recurring misconception is that all subjective-experience memory is synonymous with conscious recollection. The literature is narrower in some places and broader in others. In neuroscience, episodic memory is tied to autonoetic consciousness and mental time travel (Pastor, 2020). In agent research, however, “subjective experience” often denotes the agent’s own action–observation trajectories, behavioral histories, or distilled procedural insights, whether or not any claim about phenomenal consciousness is made (Zhang et al., 2024). A second misconception concerns terminology: in RMU, “reinforcement memory” is derived from behavioral psychology and “does not employ reinforcement learning (RL) algorithms; there are no rewards, policies, or value functions” (Hu et al., 2022).

Several substantive debates remain unresolved. The multiple-memory-systems review identifies continuing disputes over single- versus dual-process recognition memory, the necessity and sufficiency of hippocampus for recollection, and the precise contribution of angular gyrus, precuneus, and medial prefrontal cortex to vividness, confidence, and self-projection (Pastor, 2020). In computational subjective-time models, the formal substrate also varies sharply: STM relies on prediction-error segmentation, the synaptic clock relies on trace persistence, and the self-simulational account defines temporal width through policy-conditioned divergence (Zakharov et al., 2020, Jura, 2018, Bellingrath, 2024). This suggests that “subjective” is not a single mechanism but a family resemblance across distinct operationalizations.

Technical limitations are equally explicit. STM may be weakened in environments with weak event structure, and fixed thresholds $h = W_e \tilde{x}, \qquad \hat{x} = W_c h,$1 can be brittle; its action summarization heuristic is acknowledged as simple but lossy (Zakharov et al., 2020). RMU has been validated only on VQA and video QoE, depends on hand-crafted input features in those experiments, and uses piecewise-linear max-based updates that may create non-smooth gradients at switching points (Hu et al., 2022). MCT leaves the dimensionality and computation of the information-density vector implementation-dependent and does not specify a formal mapping from vector amplitude to memory write strength (Gillon, 2 Oct 2025). LLM-agent memory systems face scalability, latency, noise accumulation, privacy and security issues, and a lack of standardized benchmarks that directly isolate memory quality from overall task performance (Zhang et al., 2024).

Future directions follow directly from these limitations. The cited work proposes dynamic thresholds and richer action summaries for subjective-timescale models, hierarchical or attention-augmented appraisal units, uncertainty- or confidence-aware retrieval, multimodal memory organization, distributed shared memory, explicit provenance and temporal grounding, and experience-stage benchmarks that evaluate abstraction rather than mere storage (Zakharov et al., 2020, Hu et al., 2022, Luo et al., 7 May 2026, Dorsey et al., 4 May 2026). Across these proposals, the central research problem remains stable: how to make memory selective, temporally grounded, and behaviorally useful without collapsing it into either raw logging or opaque parameterization.