Enhanced Mycelium of Thought (EMoT)

• Enhanced Mycelium of Thought (EMoT) is a bio-inspired reasoning architecture that organizes processing into a four-level hierarchy with strategic dormancy to preserve low-scoring insights.
• It integrates a Memory Palace featuring five mnemonic encoding styles to maintain persistent, retrievable intermediate insights across multi-domain reasoning tasks.
• EMoT draws on fungal computing research by leveraging mycelial network analogies to model distributed, adaptive, and persistent pathways for complex problem solving.

Enhanced Mycelium of Thought (EMoT) denotes a bio-inspired reasoning architecture for LLMs that organises cognitive processing into a four-level hierarchy (Micro, Meso, Macro, Meta), implements strategic dormancy and reactivation of reasoning nodes, and integrates a Memory Palace with five mnemonic encoding styles; it is presented as a research prototype for complex, multi-domain problems rather than a general-purpose prompting enhancement (Stummer, 25 Mar 2026). The designation also carries a literal mycelial lineage. Earlier work on fungal computing treated living mycelium and mycelium-bound composites as substrates in which electrical propagation, nonlinear transformation, and morphology-dependent circuitry can realise Boolean logic, thereby supplying the biological analogy of distributed, redundant, adaptive pathways that the later LLM architecture formalises in software (Beasley et al., 2021).

1. Conceptual scope and intellectual lineage

EMoT was introduced to address limitations attributed to Chain-of-Thought (CoT) and Tree-of-Thoughts (ToT). CoT is characterised as linear, lacking backtracking, parallel hypotheses, and persistent state across iterations. ToT is described as branching but static: once a branch is pruned, it is gone permanently, so a weak early hypothesis cannot be reconsidered later if new evidence makes it relevant. Neither approach is said to provide persistent memory as a structured, retrievable repository of intermediate insights that survives across reasoning passes, and neither models strategic dormancy, namely the preservation of a discarded thought in reserve rather than its deletion (Stummer, 25 Mar 2026).

The biological analogy is not incidental. The motivating image is a mycelial network that is decentralized, multi-scale, redundant, and capable of keeping pathways available for later reactivation. The framework is tied particularly to clinical and multi-system reasoning, where initial hypotheses may require reconsideration as evidence evolves. The paper explicitly relates this to Croskerry’s dual-process account of diagnostic reasoning, in which an initial fast “Type 1” intuition may prematurely close alternatives while “Type 2” analysis reopens and tests them (Stummer, 25 Mar 2026).

A broader conceptual ancestry appears in fungal-computing research. That literature describes living mycelium networks as capable of efficient sensorial fusion over very large areas and distributed decision making, with information processing implemented through propagation of electrical and chemical signals together with morphological changes in the mycelium structure. This suggests that the phrase “Enhanced Mycelium of Thought” functions both as a literal biological metaphor and as a programmatic claim that distributed substrate intelligence can be abstracted into an engineered reasoning architecture (Adamatzky et al., 2021).

2. Mycelial computation as the biological precedent

A central antecedent to EMoT is the modelling of a real fungal colony as a computable electrical substrate. In one study, a 3D fluorescent Z-stack of Aspergillus niger was converted into a weighted Euclidean graph: vertices corresponded to junctions or endpoints, edges to hyphal segments, and edge weights to Euclidean distances between connected points. That graph was then mapped to resistive-capacitive networks in which the electrical parameters of edges were functions of edge length, with resistances in the kΩ range and capacitances in the pF range. In compact form, each edge $e\in E$ with length $\ell_e$ was associated with

$R_e = R(\ell_e), \qquad C_e = C(\ell_e).$

The model was analysed in SPICE using transient simulations with low-voltage pulses representing Boolean inputs $0$ and $1$; simulations ran for $40$ s with $1$ ms time steps (Beasley et al., 2021).

The computational significance of this pipeline lies in the fact that the colony’s geometry becomes part of the computation. The result is not a generic RC circuit but a geometry-derived circuit. Because the network is branching, irregular, heterogeneous, and path-asymmetric, different input sites can couple to an output site through different combinations of resistive and capacitive pathways. This was proposed as a route toward prototyping electrical analog computers from living mycelium networks, including networks hybridised with nanoparticles, and as a possible basis for smart living structures that sense cues such as light, chemicals, gases, gravity, and electric fields (Beasley et al., 2021).

A second line of modelling employed excitable-media dynamics. A projected fungal colony image on a $1000 \times 960$ grid was simulated with FitzHugh–Nagumo dynamics using the equations

$$\frac{\partial v}{\partial t} = c_1 u (u-a)(1-u) - c_2 u v + I + D_u \nabla^2$$

and

$\frac{\partial v}{\partial t} = b (u-v).$

The text notes a typographical issue in the variable naming, but the intended interpretation is an excitable-wave model with $\ell_e$0 as trans-membrane potential, $\ell_e$1 as the slow recovery variable, $\ell_e$2 as external stimulation current, and $\ell_e$3 as spatial coupling. Electrode readings were approximated by

$\ell_e$4

This line of work treated the mycelium not merely as conductive matter but as an active excitable substrate in which spike timing and wave interaction can encode logic (Adamatzky et al., 2021).

3. Logical circuits in living fungal materials

The strongest empirical basis for the mycelial analogy came from laboratory mining of logical circuits in mycelium-bound composites. In that work, a hemp shavings composite colonized by grey oyster fungus, Pleurotus ostreatus, was maintained at $\ell_e$5, relative humidity $\ell_e$6, and substrate humidity about $\ell_e$7. The hardware comprised a laptop or PC, Arduino Mega 2560, AD9833 programmable signal generators, four platinum input electrodes, seven differential DAQ output channels, and an ADC connected to a Pico 24 unit. The platinum input rods had diameter $\ell_e$8 mm, insertion depth $\ell_e$9 mm, and $R_e = R(\ell_e), \qquad C_e = C(\ell_e).$0 mm spacing in a straight line; the seven output probes were placed in a parallel line $R_e = R(\ell_e), \qquad C_e = C(\ell_e).$1 mm away and separated by $R_e = R(\ell_e), \qquad C_e = C(\ell_e).$2 mm. The experiment was repeated $R_e = R(\ell_e), \qquad C_e = C(\ell_e).$3 times, reusing the same substrate, with water spraying and a one-hour rest between repeats (Roberts et al., 2021).

Boolean inputs were encoded by amplitude,

$R_e = R(\ell_e), \qquad C_e = C(\ell_e).$4

and all $R_e = R(\ell_e), \qquad C_e = C(\ell_e).$5-bit states from $R_e = R(\ell_e), \qquad C_e = C(\ell_e).$6 to $R_e = R(\ell_e), \qquad C_e = C(\ell_e).$7 were presented in binary count order, each state lasting one hour. Outputs were decoded from peaks in the DAQ channels. Samples were recorded at $R_e = R(\ell_e), \qquad C_e = C(\ell_e).$8 Hz, peaks were detected under $R_e = R(\ell_e), \qquad C_e = C(\ell_e).$9 thresholds from $0$0 mV to $0$1 mV in steps of $0$2 mV, logical $0$3 was assigned if a peak exceeded the threshold band, otherwise $0$4, and polarity was ignored. For a given channel and threshold,

$0$5

This yielded $0$6 individual $0$7-input, $0$8-output Boolean functions, together with $0$9 state graphs, $1$0 truth tables, and $1$1 unique Boolean functions in the supplementary material (Roberts et al., 2021).

The observed function family ranged from trivial constants to standard gates and mixed sum-of-products (SOP) expressions. The two most frequent functions were False $1$2 and True $1$3. The most frequent non-trivial function was

$1$4

a $1$5-input NAND, with count $1$6. Other standard gates included

$1$7

for OR with count $1$8, and

$1$9

for AND with count $40$0. More complex examples included

$40$1

and

$40$2

The same study classified the discovered functions by the complexity of one-dimensional cellular automata governed by those functions, using

$40$3

and reported representative functions from all classes of cellular automata complexity, including the computationally universal (Roberts et al., 2021).

A synthesis paper placed these laboratory results beside RC-network simulations and spike-based gate extraction. It reported that living mycelium networks can implement select-$40$4, select-$40$5, $40$6, $40$7, $40$8, $40$9, and $1$0, while passive RC abstractions were more limited: serial RC networks produced AND, select, and AND-NOT but no OR or XOR; parallel RC networks produced AND, OR, and select but no XOR. The same paper emphasised that extracellular voltage spikes are slow, with minimum duration about $1$1 minutes and up to $1$2 hour, which constrains high-speed digital interpretation but reinforces the picture of slow, distributed, biologically embedded computation (Adamatzky et al., 2021).

4. Hierarchical reasoning topology in the 2026 EMoT architecture

The 2026 EMoT framework transposes the mycelial analogy into a fixed four-level hierarchy. The Micro level handles atomic details such as facts, observations, measurements, and single evidence items. The Meso level detects patterns across Micro outputs and performs correlation detection, temporal sequencing, and hypothesis formation. The Macro level synthesizes Meso patterns into coherent solution candidates such as diagnoses, policies, or strategic recommendations. The Meta level provides oversight and strategy by evaluating coherence, identifying gaps, and directing attention to weak or underexplored areas (Stummer, 25 Mar 2026).

Information flow is explicitly bidirectional. Bottom-up flow proceeds Micro $1$3 Meso $1$4 Macro $1$5 Meta, so that small observations aggregate into patterns and then into higher-order solution hypotheses. Top-down flow proceeds Meta $1$6 Macro $1$7 Meso $1$8 Micro, allowing strategic constraints, priorities, and correction signals to refine lower-level processing. Lateral flow also exists within each level, enabling cross-domain matching and synthesis among nodes at the same level. This differs from CoT’s single chain and ToT’s branch-only search because reasoning is treated as a network of interacting levels rather than a path or tree (Stummer, 25 Mar 2026).

The framework also specifies a set of named enhancement modules: QAM (Quality Amplification Module), IDE (Insight Distillation Engine), CEO (Computational Efficiency Optimiser), HIF (Hierarchical Integration Framework), and SDC (Strategic Dormancy Controller). The implementation is described as a Python system of $1$9 lines for the main module and about $1000 \times 960$0 lines total across $1000 \times 960$1 source files, with $1000 \times 960$2 regression tests passed. It uses NumPy, NetworkX, and scikit-learn, specifically DBSCAN and TF-IDF for clustering in the IDE, and supports pluggable backends including Anthropic Claude, Google Gemini, Ollama or local models, and a deterministic stub for testing (Stummer, 25 Mar 2026).

This architecture is presented not as a prompt template but as a reasoning scaffold. A plausible implication is that the “mycelium” in EMoT is operationalised as topology, persistence, and cross-scale coordination rather than as merely decorative biomorphic language.

5. Strategic dormancy, mnemonic encoding, and the reasoning cycle

Strategic dormancy is the framework’s most distinctive mechanism. Instead of deleting low-scoring thoughts, EMoT moves them into dormancy, where they are preserved, annotated with context, and made available for later reactivation if the reasoning environment changes. The operational trigger is a node-level Trust Score,

$1000 \times 960$3

where $1000 \times 960$4 is Success Likelihood, $1000 \times 960$5 is Novelty, $1000 \times 960$6 is Depth, and $1000 \times 960$7 is Coherence. If a node’s Trust Score falls below the heuristic threshold $1000 \times 960$8, it is moved to dormancy rather than discarded. The paper explicitly states that the weighting scheme and threshold are heuristic and were not tuned by systematic hyperparameter search (Stummer, 25 Mar 2026).

The Strategic Dormancy Controller uses three mechanisms. Predictive Relevance Modelling retains metadata about a dormant node’s content, creation context, and predicted conditions under which it may matter again. A Partial Activation Mechanism periodically reevaluates dormant nodes against the current context and may allow reduced-weight contribution before full reactivation. A Temporal Reasoning Engine tracks the evolution of the task over time and searches for phase transitions, for example from hypothesis generation to hypothesis testing, when a previously weak idea may become important. The framework does not report detailed statistics on reactivation frequency or on the contribution of reactivated nodes to the final output (Stummer, 25 Mar 2026).

Persistent memory is provided by a Memory Palace inspired by the classical method of loci. The paper defines five encoding styles: Visual Hook, Loci Room, Chunking, Temporal Ladder, and Narrative Hook. Visual Hook stores an insight as a vivid image-like representation; Loci Room stores insights spatially in a virtual room; Chunking compresses multiple related insights into a smaller retrievable unit; Temporal Ladder orders insights by time; Narrative Hook embeds insights inside a causal story. The authors do not claim that all five are equally powerful in every setting. Instead, they suggest that clinical reasoning may benefit from Temporal Ladder and Chunking, policy reasoning from Narrative Hook, and spatial relational tasks from Loci Room (Stummer, 25 Mar 2026).

The end-to-end workflow is described in eight stages: problem ingestion, perception, processing, propagation, trust evaluation, dormancy or reactivation cycle, memory encoding, and distillation and integration. During processing, a node may generate text directly, call an LLM through the pluggable backend, or apply rule-based logic. Outputs propagate upward, downward, and laterally depending on level and relevance, after which the final answer is synthesised at higher levels. This suggests that EMoT is designed as a recurrent reasoning substrate rather than a one-pass prompting pattern (Stummer, 25 Mar 2026).

6. Evaluation, performance profile, and limitations

EMoT was evaluated in two main ways. The first was a blind LLM-as-Judge benchmark using Claude Sonnet 4, with method labels removed and three independent runs. The rubric had six criteria scored from $1000 \times 960$9 to $$\frac{\partial v}{\partial t} = c_1 u (u-a)(1-u) - c_2 u v + I + D_u \nabla^2$$0: Recursion Depth, Dormant Thought Management, Cross-Domain Synthesis, Memory Utilisation, Structured Output, and Solution Quality. The second was a $$\frac{\partial v}{\partial t} = c_1 u (u-a)(1-u) - c_2 u v + I + D_u \nabla^2$$1-item short-answer benchmark across five categories—mathematical reasoning, logical reasoning, multi-step QA, planning, and BBH tasks—compared against Direct prompting, CoT, Self-Consistency, and EMoT (Stummer, 25 Mar 2026).

In the blind LLM-as-Judge benchmark, overall mean scores were EMoT $$\frac{\partial v}{\partial t} = c_1 u (u-a)(1-u) - c_2 u v + I + D_u \nabla^2$$2 and CoT $$\frac{\partial v}{\partial t} = c_1 u (u-a)(1-u) - c_2 u v + I + D_u \nabla^2$$3. Run-to-run stability was EMoT SD $$\frac{\partial v}{\partial t} = c_1 u (u-a)(1-u) - c_2 u v + I + D_u \nabla^2$$4 and CoT SD $$\frac{\partial v}{\partial t} = c_1 u (u-a)(1-u) - c_2 u v + I + D_u \nabla^2$$5. Per-criterion means were: Recursion Depth, EMoT $$\frac{\partial v}{\partial t} = c_1 u (u-a)(1-u) - c_2 u v + I + D_u \nabla^2$$6 vs CoT $$\frac{\partial v}{\partial t} = c_1 u (u-a)(1-u) - c_2 u v + I + D_u \nabla^2$$7; Dormant Thought Management, EMoT $$\frac{\partial v}{\partial t} = c_1 u (u-a)(1-u) - c_2 u v + I + D_u \nabla^2$$8 vs CoT $$\frac{\partial v}{\partial t} = c_1 u (u-a)(1-u) - c_2 u v + I + D_u \nabla^2$$9; Cross-Domain Synthesis, EMoT $\frac{\partial v}{\partial t} = b (u-v).$0 vs CoT $\frac{\partial v}{\partial t} = b (u-v).$1; Memory Utilisation, EMoT $\frac{\partial v}{\partial t} = b (u-v).$2 vs CoT $\frac{\partial v}{\partial t} = b (u-v).$3; Structured Output, $\frac{\partial v}{\partial t} = b (u-v).$4 vs $\frac{\partial v}{\partial t} = b (u-v).$5; Solution Quality, EMoT $\frac{\partial v}{\partial t} = b (u-v).$6 vs CoT $\frac{\partial v}{\partial t} = b (u-v).$7. The paper therefore characterises Cross-Domain Synthesis as EMoT’s standout strength while acknowledging that CoT is slightly better on most other criteria. In the Patient Bengt case, EMoT averaged $\frac{\partial v}{\partial t} = b (u-v).$8 and CoT $\frac{\partial v}{\partial t} = b (u-v).$9, in a setting requiring integration across haematology, geriatrics, endocrinology, pharmacology, and supply chain considerations (Stummer, 25 Mar 2026).

The ablation results are central to the framework’s internal argument. Full EMoT scored $\ell_e$00; EMoT with no dormancy scored $\ell_e$01; EMoT with no Memory Palace scored $\ell_e$02. The paper interprets the no-dormancy result as collapse, arguing that without SDC the efficiency and pruning machinery eliminates all low-trust nodes and prevents meaningful synthesis. By contrast, the memory system contributes a modest but measurable benefit, especially on Memory Utilisation and Cross-Domain Synthesis (Stummer, 25 Mar 2026).

On the short-answer benchmark, the performance profile reverses. Direct prompting achieved $\ell_e$03, CoT $\ell_e$04, Self-Consistency $\ell_e$05, and EMoT $\ell_e$06. The paper identifies this as evidence of systematic overthinking on simple problems. Computational cost is also substantial: on the three-case complex benchmark EMoT used $\ell_e$07 LLM calls versus $\ell_e$08 for CoT, about $\ell_e$09 more calls, with token overhead about $\ell_e$10, runtime about $\ell_e$11, and cost estimate about $\ell_e$12. Methodological cautions are explicit: small sample sizes $\ell_e$13 complex cases, $\ell_e$14 short-answer items$\ell_e$15, LLM-as-Judge evaluation with potential self-preference bias, rubric circularity, no human expert validation for the clinical case, no systematic tuning of trust weights or dormancy threshold, and no fine-grained reactivation analysis. The clinical case is synthetic and illustrative only, and the framework is stated not to be a clinical decision support system and not to have been clinically validated (Stummer, 25 Mar 2026).

A recurrent misconception is therefore addressed directly in the source: EMoT is not presented as a universal reasoning booster. The authors argue that CoT may be enough, or better, for predictable and structured problems, whereas EMoT may be beneficial when the problem is multi-domain, hypothesis revision is necessary, premature pruning is dangerous, or information is hidden or evolving (Stummer, 25 Mar 2026).

7. Terminological ambiguity and adjacent uses of the acronym

The acronym “EMoT” is not unique to the reasoning framework. In quantitative finance, “EMoT” denotes Entropic Martingale Optimal Transport, and “c-EMOT” denotes a constrained or certificate-bearing implementation of that idea. In the SPX–VIX surface-construction setting, c-EMOT is a tri-marginal, martingale-constrained entropic OT bridge solved numerically with certificates, positioned between constructive PCA–Smolyak approximation and weighted projection onto the no-arbitrage cone (Zhang, 12 Nov 2025).

In that usage, the core formulation is

$\ell_e$16

subject to

$\ell_e$17

with a log-domain tri-Sinkhorn solver, low-rank kernel surrogates, spectral whitening, $\ell_e$18-annealing, adaptive damping, and certificates such as the KKT residual, geometric decay ratio $\ell_e$19, and strong-convexity lower bound $\ell_e$20. The paper reports pass values including $\ell_e$21 in one macro list, $\ell_e$22 with threshold $\ell_e$23, empirical Lipschitz $\ell_e$24, dupok=True, and total risk $\ell_e$25 in the macro definitions (Zhang, 12 Nov 2025).

This second usage is terminologically adjacent but conceptually separate from Enhanced Mycelium of Thought. One refers to an LLM reasoning architecture grounded in a mycelial metaphor and fungal-computing antecedents; the other refers to a certified entropic martingale transport bridge in mathematical finance. The shared acronym therefore requires contextual disambiguation.