Endogenous Recall & Task-Splitting

• Endogenous recall is the spontaneous internal retrieval of stored sub-task solutions that enables rapid adaptation and efficient memory reuse in both biological and artificial systems.
• Task-splitting decomposes complex problems into distinct, manageable sub-tasks, as seen in transformer attention circuits and associative memory models.
• Together, these phenomena reduce interference, enhance performance in continual learning, and support scalable, modular architectures for adaptive task management.

Endogenous recall is the spontaneous, model-internal retrieval of stored content or sub-task solutions without external cuing, often accompanying the decomposition of complex behaviors into separable sub-tasks—a process referred to as task-splitting. This dual phenomenon is central in diverse domains, from associative memory models and continual learning systems to transformer networks and LLM task automation. Task-splitting reveals how learning systems allocate memory and computation across temporally or structurally distinct sub-tasks, and endogenous recall underpins the reuse and rapid adaptation that characterize skilled biological and artificial agents.

1. Formal Mechanisms of Endogenous Recall

Endogenous recall manifests through a spectrum of algorithmic and representational mechanisms, from discrete associative memory to context-conditioned policy reinstatement. In classical associative networks, such as the Blume–Emery–Griffiths (BEG) neural network, recall involves the spontaneous reactivation of stored patterns upon suitable input cues, requiring no external index beyond the stored representation itself (Albanese et al., 12 Jan 2026). In meta-learning with episodic reinstatement architectures, a differentiable neural dictionary (DND) stores the internal state of an agent at episode termination, and gating mechanisms merge retrieved memories directly into working memory when a familiar context returns (Ritter et al., 2018). Similarly, in Automatic Recall Machines (ARM), endogenous replay samples are generated by optimizing interference metrics over the model’s own prediction graph, with no buffer or generator external to the main model (Ji et al., 2020).

In the context of transformer architectures, endogenous recall can emerge as specialized attention circuits that retrieve and apply previously encountered system states or label mappings, triggered by in-context queries (e.g., associative recall by symbolic label, or Bayesian-style continuation by state similarity) (Daniels et al., 2 Jul 2025). In application-oriented systems such as MobileGPT, endogenous recall is realized by classifying current app screens into previously encountered “page nodes” using internal key-UI patterns, then replaying or adapting stored sub-task routines without re-derivation (Lee et al., 2023).

2. Task-Splitting: Structural and Emergent Decomposition

Task-splitting denotes the internal segmentation of a nominally unified problem into distinct cognitive, computational, or representational sub-parts. This splitting may be designed (as in hierarchical RL or macroaction learning), contextually gated (as in LLM-based modular controllers), or may emerge endogenously as a result of sufficient task complexity or memory constraints.

In BEG associative memory, the sparsity of stored patterns (i.e., high dilution) triggers a qualitative transition from serial recall (recall of one pattern at a time) to parallel recall (simultaneous recall of multiple patterns), with intermediate regimes displaying hierarchical recall amplitudes (Albanese et al., 12 Jan 2026). In in-context learning with transformers, task-splitting is evidenced by the distinct phase dynamics of label-based associative recall (required for “initiation” of a resumed sub-task) and observation-based Bayesian continuation (for continuing sequences)—with abrupt emergence of associative recall and more gradual improvement in continuation performance (Daniels et al., 2 Jul 2025).

Table: Emergence of Task-Splitting Phenomena

Domain/Model	Subtask Types	Evidence
BEG associative net	Pattern segments	Serial→parallel regimes
Transformer ICL tasks	Recall/initiation vs continuation	Distinct phase transitions
MobileGPT task automation	Sub-tasks per app page	Hierarchical recall/adapt
ARM continual learning	Class-based micro-tasks	Entropy-driven splitting

Task-splitting is not always externally labeled; it may express itself through architectural specialization, lateralization (e.g., memory banks with inhibitory cross-talk (Jeong, 27 Feb 2026)), or via optimization pressures that favor distributed over saturated representations.

3. Architectural and Algorithmic Substrates

A range of architectures implement endogenous recall with task-splitting via explicit and implicit currents.

• Associative memory networks: In the BEG model, $\sigma_i \in \{-1, 0, +1\}$ neurons store diluted patterns $\xi^\mu$, and recall arises from the minimization of the Hebbian energy given partial input. The system naturally splits retrieval across available neurons when dilution is high, creating graded parallel recall where each “slot” may serve a different stored pattern (Albanese et al., 12 Jan 2026).
• Meta-learning with episodic memory (epLSTM): The introduction of a reinstatement gate ($r_t$) merging retrieved cell states into the standard LSTM core enables rapid reinstatement of prior behavior within re-encountered tasks, providing selective, context-driven, endogenous recall (Ritter et al., 2018).
• Memory-augmented transformers: Models with physically partitioned memory banks and sign-controlled cross-talk matrices achieve sharp functional task-splitting: persistent episodic memory is localized to one bank via active inhibitory suppression of the contralateral bank, supporting non-interfering recall for associative tasks while enabling rule extraction in the complementary bank (Jeong, 27 Feb 2026).
• LLM-powered procedural automation: MobileGPT’s hierarchical task graph $(P, E, T)$, recall-by-key-UI pattern matching, and exact/fallback gating constitute a human-like, hierarchical task-splitting substrate, permitting efficient reuse via sub-task recall and dynamic in-context adaptation when memory fails (Lee et al., 2023).

4. Training Dynamics and Quantitative Phase Transitions

Emergence of endogenous recall and intrinsic task-splitting is tightly coupled to training dynamics and phase transitions.

• Transformer next-token prediction: Distinct capabilities—the continuation of a resumed sequence and the associative, label-based recall required to initiate resumption—show sharply different learning curves. Bayesian continuation emerges early and smoothly, while associative recall emerges later and abruptly, as demonstrated by median MSE curves in the “needle-in-a-haystack” tests (with ≈2×10⁶ steps for basic continuation, ≈1.5-2.5×10⁷ for associative recall, and convergence of continuation by ≈6×10⁷ steps) (Daniels et al., 2 Jul 2025).
• Functional lateralization: In memory-augmented transformers, specialization metrics $\mathcal D_{sep}$ and $\mathcal P_{ct}$ quantify endogenous bank specialization. Inhibitory cross-talk yields $\mathcal D_{sep} = \pm 1$ (maximal specialization), $\mathcal P_{ct} \approx 0$ (no cross-task confusion), corresponding to 124-fold improvement in cipher recall loss compared to baseline (Jeong, 27 Feb 2026).
• Episodic meta-recall: Analysis of $r$-gate activations in epLSTM reveals conditional opening on cued trials and when recalled episodes are exploited, indicative of learning to segment and recall without explicit supervisory signals (Ritter et al., 2018).
• Continual learning recall: ARM methods show that batch-specialized recall, enforced by diversity penalties in the recall objective, yields low forgetting (e.g., ~8% vs ~50% with naive SGD) while using less memory than buffer-based experience replay (Ji et al., 2020).

5. Out-of-Distribution Probes and Mechanistic Decomposition

Mechanistic analyses reveal the separability and context-dependence of recall and continuation mechanisms.

• Interventional OOD tasks: Swapping, synchronizing, or misdirecting in-context labels in transformer ICL tasks selectively disables or preserves only the sub-component of the learned mechanism (e.g., breaking associative recall but leaving observation-based continuation untouched), evidencing the existence of distinct subcircuits (Daniels et al., 2 Jul 2025).
• Circuit disjointness via pruning: Edge-pruning analysis demonstrates that the “1-after” (label/gate-based recall) and “2-after” (label-agnostic continuation) mechanisms correspond to disjoint (0% edge overlap) sparse subcircuits in late-checkpoint models (Daniels et al., 2 Jul 2025).
• Parallel recall in diluted associative nets: As pattern dilution increases, BEG networks naturally distribute retrieval effort, leading to hierarchical recall (in which the leading pattern dominates but subleading patterns are retrieved proportionally to available free neurons) or, at higher dilution, to equal-strength parallel recall (Albanese et al., 12 Jan 2026). Analytically, the crossover occurs at $a_c = \frac{1}{2}$ for large $\xi^\mu$0, separating the hierarchical and fully parallel regimes.

6. Cognitive, Theoretical, and Applied Implications

The emergence of endogenous recall and task-splitting is theoretically significant for several domains:

• Biological memory: Sparse neural coding and graded response enable distributed retrieval in neural circuits, paralleling the phase behavior observed in associative memory networks (Albanese et al., 12 Jan 2026). Top-down modulation in ARM models aligns with feedback and replay processes in hippocampal-cortical systems (Ji et al., 2020).
• Flexibility and Adaptation: The ability to decompose tasks and recall sub-task routines underpins the modularity, compositionality, and rapid adaptation seen in human skills and advanced LLM-driven systems. MobileGPT, for instance, achieves adaptive task re-use, lowering latency and operational costs by 62.5% and 68.8%, respectively, compared to non-memory baselines (Lee et al., 2023).
• Catastrophic interference and continual learning: Endogenous, batch-specialized recall mechanisms dramatically lower forgetting without the need for external data buffers or generative replay, offering a scalable path for continual learning agents (Ji et al., 2020).

A plausible implication is that architectures supporting endogenous recall with flexible, context-gated task-splitting will possess scalable, robust memory reuse and reduced interference—key for both biological intelligence and practical large-scale AI systems.

7. Open Problems and Future Directions

While the mechanistic bases for endogenous recall and task-splitting are increasingly clarified, open challenges remain:

• Unsupervised segmentation: Real-world tasks often lack discrete episode boundaries; learning to detect and encode salient events for episodic storage is unresolved in both LSTM-based meta-learning and memory-augmented transformers (Ritter et al., 2018).
• Granularity of task splits: Understanding how granularity is selected—whether at the level of states, labels, sub-task boundaries, or memory banks—remains an open problem, as does automating this selection without external annotation.
• Interference suppression: Further analysis of lateralization, inhibitory cross-talk, and specialized routing is needed to generalize these findings from symbolic tasks to rich multimodal settings (Jeong, 27 Feb 2026).
• Integration with planning and reasoning: How endogenous recall mechanisms can be extended to structured reasoning, hierarchical planning, or multi-agent cooperation with dynamically evolving sets of sub-tasks is largely unexplored.

These directions suggest that endogenous recall and task-splitting are central not only for understanding neural computation and artificial memory, but also for advancing modular, adaptive, and robust general intelligence across domains.