---
title: Forgetful Gauges in Electrodynamics & Memory
url: https://www.emergentmind.com/topics/forgetful-gauges
type: topic
---

# Forgetful Gauges in Electrodynamics & Memory

“Forgetful Gauges” (*Editor’s term*) designates two distinct technical uses of gauge-like constructs in the supplied literature. In classical electrodynamics, gauges are connections between the electromagnetic potentials, and Onoochin argues that the Coulomb and velocity gauges permit instantaneous or superluminal ingredients to survive in computed electric fields, so that only the Lorenz gauge should be used in applied calculations [2510.08583]. In information systems, by contrast, Memory Buoyancy is explicitly presented as a dynamic “forgetful gauge” that measures each information item’s current value and drives Managed Forgetting in a Semantic Desktop [1811.12177]. A related operational perspective appears in agent-memory evaluation, where ForgetEval studies how control-plane placement determines which forgetting failure modes a system can recover across thirteen system configurations [2606.15903]. The commonality is not a shared formalism but a shared role: a gauge specifies what is retained, what decays, and under what constraints.

## 1. Terminological scope and conceptual role

In the electrodynamic usage, a gauge is a relation imposed on the four-potential
$$
A^\mu = (\phi,\vec A),
$$
with fields recovered as
$$
\vec E = -\nabla \phi - \frac{1}{c}\partial_t \vec A,\qquad
\vec B = \nabla \times \vec A.
$$
The supplied material identifies three gauges that can be used to describe systems of charges and currents without restrictions: the Lorenz gauge, the Coulomb gauge, and the velocity gauge. It also states that more specific gauges are reductions of the Lorenz gauge [2510.08583].

In the Semantic Desktop literature, the “gauge” is explicitly metaphorical. Memory Buoyancy assigns every “thing” in a Personal Information Model a normalized score $MB\in[0,1]$ indicating its current value for a particular user. That score then drives an escalating spectrum of forgetting actions, from temporary hiding through condensation and adaptive synchronization to archiving or deletion [1811.12177].

ForgetEval uses neither the electrodynamic nor the Memory Buoyancy terminology, but it introduces a third gauge-like mechanism: deterministic scoring of whether a memory system forgets correctly after control-plane mutations. This suggests a broader pattern in which a “forgetful gauge” is a device that makes retention and deletion operationally testable. That implication is interpretive rather than terminological, but it is consistent with the supplied descriptions of Memory Buoyancy and ForgetEval [2606.15903].

## 2. Electrodynamic formulation: Lorenz, Coulomb, and velocity gauges

The Lorenz gauge is defined by
$$
\partial_\mu A^\mu = 0,
$$
or, in three-vector form,
$$
\nabla\cdot \vec A_L + \frac{1}{c}\partial_t \phi_L = 0.
$$
Its potentials are retarded at the light speed $c$:
$$
\phi_L(\vec r,t)=\frac{1}{4\pi}\int \frac{\rho(\vec r',\,t-|\vec r-\vec r'|/c)}{|\vec r-\vec r'|}\,d^3r',
$$
$$
\vec A_L(\vec r,t)=\frac{1}{4\pi c}\int \frac{\vec J(\vec r',\,t-|\vec r-\vec r'|/c)}{|\vec r-\vec r'|}\,d^3r'.
$$
The fields are then
$$
\vec E_L=-\nabla\phi_L-\frac{1}{c}\partial_t\vec A_L,\qquad
\vec B_L=\nabla\times\vec A_L.
$$
The supplied material characterizes these fields as manifestly causal and propagating at $c$ [2510.08583].

The Coulomb gauge is defined by
$$
\nabla\cdot \vec A_C=0.
$$
Its scalar potential is the instantaneous solution of Poisson’s equation,
$$
\nabla^2\phi_C(\vec r,t)=-4\pi \rho(\vec r,t),
$$
so that
$$
\phi_C(\vec r,t)=\frac{1}{4\pi}\int \frac{\rho(\vec r',t)}{|\vec r-\vec r'|}\,d^3r'.
$$
The vector potential satisfies
$$
\left(\nabla^2-\frac{1}{c^2}\partial_t^2\right)\vec A_C
= -\frac{4\pi}{c}\vec J-\frac{1}{c}\nabla(\partial_t\phi_C),
$$
and the fields are computed from the same reconstruction formulas,
$$
\vec E_C=-\nabla\phi_C-\frac{1}{c}\partial_t\vec A_C,\qquad
\vec B_C=\nabla\times\vec A_C.
$$

The velocity gauge is a one-parameter family labeled by $v$ and defined by
$$
\nabla\cdot \vec A_v+\frac{c}{v^2}\partial_t\phi_v=0.
$$
This yields the decoupled equations
$$
\left(\nabla^2-\frac{1}{v^2}\partial_t^2\right)\phi_v=-4\pi \rho,
$$
$$
\left(\nabla^2-\frac{1}{c^2}\partial_t^2\right)\vec A_v
= -\frac{4\pi}{c}\vec J
-c\left(\frac{1}{v^2}-\frac{1}{c^2}\right)\nabla(\partial_t\phi_v).
$$
The scalar potential is retarded at speed $v$,
$$
\phi_v(\vec r,t)=\frac{1}{4\pi}\int \frac{\rho(\vec r',\,t-|\vec r-\vec r'|/v)}{|\vec r-\vec r'|}\,d^3r',
$$
while $\vec A_v$ is obtained by the usual $c$-retarded Green’s function acting on the source term above. The limiting relations are explicit: $v=c$ recovers the Lorenz gauge, and $v\to\infty$ recovers the Coulomb gauge [2510.08583].

## 3. Causality critique and the claim that Coulomb and velocity gauges are unusable in applications

The central controversy in Onoochin’s treatment is that, although it is commonly accepted that the Lorenz, Coulomb, and velocity gauges are equivalent in the sense that they yield identical electromagnetic fields, the Coulomb and velocity gauges are said to produce solutions corresponding to superluminal propagation of the electric field [2510.08583]. The paper’s conclusion is categorical: because such propagation has not been observed experimentally and is forbidden by special relativity, calculations in the Coulomb and velocity gauges may yield incorrect results and therefore cannot be used in applied electromagnetic calculations.

For the Coulomb gauge, the scalar potential is described as an instantaneous “action-at-a-distance” functional of $\rho(\vec r,t)$ via the Poisson Green’s function
$$
G_C(\vec r-\vec r',t-t')=\frac{\delta(t-t')}{|\vec r-\vec r'|}.
$$
Hence $\nabla\phi_C$ carries information instantly across space. The source term $\nabla(\partial_t\phi_C)$ in the wave equation for $\vec A_C$ is likewise non-local and instantaneous, and the supplied account states that the net effect is that both pieces in $\vec E_C$ contain parts that depend on $\rho$ and $\vec J$ at the same time $t$ at arbitrarily large separation [2510.08583].

A concrete example is given for a point charge $q$ at rest until $t=0$ at a point $P$ a distance $D$ from a detector $O$, after which it suddenly moves with velocity $u\ll c$. For $D>ct$, the light front from $t=0$ has not reached $O$, so a causal electric field should vanish. In the Coulomb gauge, however,
$$
E_{C,x}(O,t)
= \frac{q}{(D-ut)^2}
-\frac{1}{c}\partial_t A_{C,x}(O,t)
= \left[\frac{q}{(D-ut)^2}\right]
\left\{1-\frac{u^2}{c^2}\ln\left[1-\left(\frac{ct}{D}\right)^2\right]\right\}\neq 0
$$
even for $D>ct$. The logarithmic term is said to arise from the volume integral of $\partial_t^2\phi_C$ over the sphere of radius $ct$, which is presented as an explicit demonstration of “superluminal” or “instantaneous” arrival [2510.08583].

For the velocity gauge with finite $v>c$, the scalar potential is retarded at a superluminal speed $v$. When $\vec A_v$ is built by integrating its nonlocal source $\nabla(\partial_t\phi_v)$ over the $c$-retarded light cone, contributions arise from space-time points satisfying
$$
\frac{|\vec r-\vec r'|}{v}+\frac{|\vec r-\vec r'|}{c}\le t,
$$
which can lie outside the true light cone $|\vec r-\vec r'|/c\le t$ if $v>c$. The supplied summary states that a power-series argument shows that the net $\vec E_v$ contains terms proportional to $u^2/c^2$ that do not cancel $\nabla\phi_v$, and hence remain nonzero for spacelike separations [2510.08583].

The paper ties this directly to special relativity: no signal or physical influence can propagate faster than $c$, and electromagnetic fields carry energy and momentum. It therefore treats the uncanceled instantaneous or $v$-retarded pieces in $\vec E_C$ and $\vec E_v$ as physically unacceptable. A common misconception addressed by the paper is that formal gauge equivalence is automatically sufficient for applied work; the argument presented is that, in these gauges, the gauge-dependent pieces no longer cancel exactly at superluminal separations, so practical calculations can acquire spurious field impulses before the true signal arrives [2510.08583].

## 4. Memory Buoyancy as a dynamic forgetful gauge

In the Semantic Desktop framework, every e-mail, file, calendar entry, contact, Web page, or annotation is represented as a “thing” in a Personal Information Model connected in a rich semantic graph. Memory Buoyancy is the core mechanism by which the Semantic Desktop continuously measures each information “thing’s” short-term relevance and thereby drives Managed Forgetting. It is explicitly described as a dynamic “forgetful gauge” that tells the system which items to keep in view and which to let sink below the user’s radar [1811.12177].

The score is normalized,
$$
MB\in[0,1],
$$
and indicates an item’s current value for a particular user. When $MB$ falls below configurable thresholds, the system may hide items from search results, demote them in sidebars, synchronize them less eagerly, or eventually delete them. The supplied material emphasizes that this avoids a rigid keep-or-delete paradigm [1811.12177].

The initial formulation separates a static, time-independent component from a dynamic, time-dependent component. The static part is
$$
\mathrm{MB}_{\mathrm{stat}}(r,u,c)=\alpha_{\mathrm{type}(r)}+\gamma\times |\mathrm{neighbors}(r)|,
$$
where $r$ is the resource, $u$ the user, $c$ the context, $\alpha_{\mathrm{type}(r)}$ a base weight by resource type, $|\mathrm{neighbors}(r)|$ the number of semantic links in the PIMO, and $\gamma$ a tuning constant. For stimulation events at times $\{t_1,\dots,t_n\}$, the saturating stimulation function is
$$
s(\Delta t)=\alpha\bigl(1-e^{-\Delta t/\beta}\bigr),
$$
and the piecewise decay is
$$
\mathrm{MB}_{\mathrm{dyn}}(\Delta)=
\begin{cases}
e^{-\Delta/\tau_1}, & \Delta\le T_{\mathrm{switch}}\\[6pt]
e^{-\Delta/\tau_2}, & \Delta>T_{\mathrm{switch}}
\end{cases}
$$
with $\tau_1\ll \tau_2$. The total initial value is
$$
\mathrm{MB}_{\mathrm{initial}}(r,t)=
\mathrm{norm}\Bigl(\mathrm{MB}_{\mathrm{stat}}(r)+\sum_{i=1}^{n}s(t-t_i)\Bigr)\times
\mathrm{MB}_{\mathrm{dyn}}(t-t_n).
$$
Implementation follows the same decomposition: the static part is updated only on graph changes, while the dynamic part is computed on demand when an application requests MB [1811.12177].

The three years of daily use reported in the supplied material yielded both success and failure cases. As a success, workshop artifacts faded gracefully: immediately after a meeting, related e-mails, tasks, and slides all showed $MB\approx 1.0$; eight months later only project people and photos still appeared; two years later only the top-level project node and frequently revisited photos remained visible. As failures, the semantic-search “MB threshold” slider was habitually dragged down to $0.0$, bypassing forgetting, and documents relevant in one context resurfaced in another because the initial model ignored user context [1811.12177].

To address these limitations, the advanced version introduced Local MB, Global MB, and Group MB. Local MB measures an item’s relevance for a user within a context, and when the user switches context the old local values are frozen and do not decay further until that context is revisited. Global MB is a context-free aggregate shielded from sudden context-switch perturbations. Group MB summarizes relevance across all users in the Group Information Model and drives shared-memory aging and archival. Each variant reuses the same static-dynamic architecture but filters stimulation events and decay behavior according to context or group membership. The supplied interpretation is that this restores coherent “golden threads” across switches [1811.12177].

## 5. ForgetEval and control-plane gauges for agent-memory forgetting

ForgetEval studies forgetting not as passive decay but as correctness of control-plane mutation. Its core claim is that where intelligence is placed in the pipeline—deterministic primitives only, an LLM at inscribe time, or an LLM at mutation time—shapes which forgetting failure modes are recovered [2606.15903].

The benchmark is deterministic and substring-match based. Each test case $i$ contains `setup_facts`, control-plane calls (`supersede`, `release`, `purge`), a final query $q_i$, a `must_contain` set $C_i$, and a `must_not_contain` set $F_i$. For top-$k$ retrieved hits
$$
R_i=\{r_{i1},\dots,r_{ik}\},
$$
the indicators are
$$
\mathrm{OK}^{+}_{i}
=\Bigl[\forall c\in C_i\ \exists r\in R_i:\;c\subseteq r\Bigr],
\qquad
\mathrm{OK}^{-}_{i}
=\Bigl[\forall f\in F_i\ \forall r\in R_i:\;f\not\subseteq r\Bigr].
$$
A case passes iff
$$
\mathrm{pass}_i=\mathrm{OK}^{+}_{i}\wedge \mathrm{OK}^{-}_{i},
$$
and over $N$ cases the accuracy is
$$
\mathrm{Acc}=\frac{1}{N}\sum_{i=1}^N \mathrm{pass}_i.
$$
The suite uses $N=1000$ on the deterministic template suite and $N=385$ on the adversarial layer. Lift from baseline $B$ to enhanced system $E$ is defined as
$$
\mathrm{Lift}(E,B)=\mathrm{Acc}_E-\mathrm{Acc}_B.
$$
An example reported in the supplied material is the $+27.8$ pt joint-placement lift on the external 77-case subset [2606.15903].

The six-method Adapter Protocol keeps recall LLM-free. A backend implements `reset`, `inscribe`, and `recall_texts`, with optional `supersede`, `release`, and `purge`. If a primitive is absent, `NotImplementedError` yields an honest “N/A”; main tables exclude N/A cases from the denominator, while a strict-denominator score also treats N/A as failures [2606.15903].

The three placement regimes have complementary behavior. Deterministic primitives are strong on lexical and temporal correctness—`substring_trap ≥ 89 %`, `temporal_qualifier = 100 %`, `negation_trap ≥ 95 %`—but weak on canonicalization, with `identifier_obfuscation ≤ 5 % (2/38)` and `cross_lingual_identifier = 0 % (0/38)`. They also fail intent-aware deletion, with `compound_fact = 0 %` and `prefix_collision` reported as `82→0 % with no LLM`. Inscribe-time LLM systems recover canonicalization entirely—`identifier_obfuscation = 100 %` and `cross_lingual_identifier = 100 %`—but still fail deletion-precision and intent-aware tasks, with `prefix_collision = 0 %` and `compound_fact = 0 %`. Mutation-time LLM hooks recover canonicalization to `92–100 %`, `prefix_collision` to approximately `79 %`, and `compound_fact` to `78–85 %`, with overall `ex-compound_fact 93.3–94.2 % (345 evaluable cases)` and `including compound_fact: 91.7–93.2 % (385 cases)` [2606.15903].

The benchmark also quantifies operational trade-offs. On the 385 adversarial cases, the mutation-time hook with DeepSeek-V3 costs approximately `$0.17`. Reported latency per case on a single CPU is `~ 74 ms` for Lethe, `~ 64 ms` for LangGraph, `~ 191 ms` for MemPalace, `~ 514 ms` for Mem0 with vector plus Qdrant cold start, and `~ 2.3 s / case` for the hooked mutation-time approach, while the recall path remains unchanged. The external 77-case subset reports `Letta only: 52.5 % pass (32/61 evaluable)` and `Letta+LLM: 80.3 % pass (49/61)`, yielding the stated `+27.8 pt` lift [2606.15903].

A misconception explicitly addressed here is that memory benchmarking is equivalent to recall benchmarking. The supplied material states that production failures are predominantly forgetting failures rather than recall failures, citing cases such as rotated credentials still being suggested and GDPR-deleted records still being retrieved. ForgetEval is presented as filling that gap by decomposing five primitive families and ten adversarial categories under deterministic scoring [2606.15903].

## 6. Comparative interpretation across the three usages

Across the supplied sources, the gauge concept serves different technical purposes. In electrodynamics it constrains the potentials from which $\vec E$ and $\vec B$ are reconstructed; in Semantic Desktop systems it measures short-term relevance and triggers forgetting actions; in agent-memory evaluation it operationalizes whether forgetting mutations behave correctly under adversarial conditions [2510.08583].

The most substantive commonality is selective suppression. In the electrodynamic account, the issue is whether gauge-dependent pieces cancel so that only causal propagation survives. In Memory Buoyancy, the issue is whether semantic items sink below visibility thresholds without destroying future recoverability. In ForgetEval, the issue is whether supersede, release, and purge remove or preserve the right records under lexical, canonicalization, and intent-aware perturbations. This suggests that “forgetful gauge” is best understood as an abstract editorial umbrella for mechanisms that regulate disappearance under constraints, rather than as a single established term spanning the three literatures.

The principal divergence is normative. Onoochin’s argument is exclusionary: Coulomb and velocity gauges are said to be pathological for real, time-dependent sources, while the Lorenz gauge alone “remembers” light-speed propagation and is therefore the only safe choice for applied electromagnetic problems [2510.08583]. Memory Buoyancy is adaptive rather than exclusionary: forgetting is a design goal, provided that it preserves trust, coherence, and context sensitivity [1811.12177]. ForgetEval is diagnostic: it does not prescribe one storage substrate, but it does show that architectural placement of LLM intelligence matters more than whether an LLM is present at all, because different placements recover different forgetting failure modes [2606.15903].

Taken together, the sources indicate that a gauge becomes “forgetful” when it mediates attenuation, suppression, or deletion in a technically consequential way. In one domain this is treated as a source of unphysical artifacts; in the others it is treated as the basis of controlled memory management. The shared vocabulary is therefore structural rather than doctrinal, and the differences in what counts as acceptable forgetting are central to the topic rather than incidental.

Source: https://www.emergentmind.com/topics/forgetful-gauges