Memory-Augmented Potential Field Theory

• Memory-Augmented Potential Field Theory is a framework that integrates past trajectory data into control by dynamically reshaping the search landscape.
• It leverages memory-based potential fields and adaptive blending coefficients to modify control objectives near trap regions and enhance exploration.
• Empirical results demonstrate improved efficiency and escape rates in robotic control, power systems, and UAV navigation compared to traditional methods.

Memory-Augmented Potential Field Theory (MAPFT) denotes a framework in which historical experience is incorporated into a potential-field-like control or decision substrate so that the effective search, planning, or control landscape is reshaped online by memory rather than fixed a priori. In its explicit formulation, MAPFT integrates historical trajectory data into stochastic optimal control by constructing memory-based potential fields that encode topological features such as local minima / traps, low-gradient plateaus, and high-curvature narrow passages / saddle-like regions, and then uses these fields to alter both the control objective and the local exploration strategy (Zheng et al., 24 Sep 2025). Related work places the same broad idea in neural field theory, kernelized reinforcement learning, and multi-agent path planning, where memory is represented respectively as persistent activity, experience particles, or spatial-temporal congestion traces (Kilpatrick et al., 2017, Chiu et al., 2022, Pertzovsky et al., 28 May 2025).

1. Definition, scope, and conceptual lineage

Within the literature, MAPFT is introduced as “a unified mathematical framework that integrates historical experience into stochastic optimal control” and “dynamically constructs memory-based potential fields that identify and encode key topological features of the state space” (Zheng et al., 24 Sep 2025). Its central claim is that conventional stochastic optimal control methods, especially sampling-based methods such as MPPI, are effectively memoryless: they optimize from the current state using the current objective and sampled rollouts, but do not retain structured experience about where they got stuck before. MAPFT addresses that limitation by converting prior failures and difficult regions into an explicit memory-dependent shaping term.

Several earlier and parallel formulations instantiate the same general motif. In the neural field model of memory-guided search, a two-layer neural field stores current position as a bump attractor and previously visited locations as a persistent activity pattern bounded by a wave front; remembered locations then bias future movement decisions, producing a “potential-field-like” search landscape generated by the agent’s own path rather than imposed externally (Kilpatrick et al., 2017). In generalized reinforcement learning, the “reinforcement field” is a kernel expansion over “experience particles” in working memory, so that policy search becomes a field query over remembered state-action-value tuples rather than a purely memoryless value-iteration procedure (Chiu et al., 2022). In lifelong multi-agent path finding, artificial potential fields are interpreted as a soft spatial-temporal memory of other agents’ recent or expected trajectories; this memory-like overlay is not helpful in one-shot MAPF, but is highly effective in LMAPF, where repeated replanning and congestion avoidance dominate performance (Pertzovsky et al., 28 May 2025).

This suggests that MAPFT is best read not as a single algorithmic family, but as a broader theoretical pattern: memory acts as a field-shaping mechanism that converts trajectory history, visited regions, or accumulated interaction evidence into a dynamic bias on future motion, action selection, or sampling.

2. Core mathematical structure

The explicit MAPFT formulation begins from the discrete-time stochastic system

$$x_{t+1}=f(x_t,u_t)+\epsilon_t,\qquad \epsilon_t\sim\mathcal N(0,\Sigma),$$

with objective

$$J(\mathbf u)=\mathbb E\!\left[\sum_{t=0}^{T-1} c(x_t,u_t)+c_T(x_T)\right].$$

Under the path-integral view, optimal control is represented as an expectation over trajectories,

$$u_t^*=\int_{\mathcal T} u_t(\tau)\,p(\tau\mid x_t)\,d\tau, \qquad p(\tau\mid x_t)=\frac{1}{Z(x_t)}\exp\!\left(-\frac{1}{\lambda}S(\tau)\right),$$

where $S(\tau)$ is trajectory cost and $\lambda>0$ controls exploration versus exploitation (Zheng et al., 24 Sep 2025).

The central MAPFT object is the memory-augmented value

$$V(x,M)=\alpha(x,M)\,V_{\text{base}}(x)+(1-\alpha(x,M))\,V_{\text{mem}}(x,M).$$

Here $V_{\text{base}}(x)$ is the original task objective, $V_{\text{mem}}(x,M)$ is the memory-induced shaping term, and $\alpha(x,M)\in[0,1]$ is an adaptive blending coefficient. Far from memorized problematic regions, $\alpha\approx 1$, so the controller behaves like the base method. Near traps or difficult regions, $$J(\mathbf u)=\mathbb E\!\left[\sum_{t=0}^{T-1} c(x_t,u_t)+c_T(x_T)\right].$$0, and memory dominates (Zheng et al., 24 Sep 2025).

Memory is represented as

$$J(\mathbf u)=\mathbb E\!\left[\sum_{t=0}^{T-1} c(x_t,u_t)+c_T(x_T)\right].$$1

where $$J(\mathbf u)=\mathbb E\!\left[\sum_{t=0}^{T-1} c(x_t,u_t)+c_T(x_T)\right].$$2 is a feature location in state space, $$J(\mathbf u)=\mathbb E\!\left[\sum_{t=0}^{T-1} c(x_t,u_t)+c_T(x_T)\right].$$3 an influence radius, $$J(\mathbf u)=\mathbb E\!\left[\sum_{t=0}^{T-1} c(x_t,u_t)+c_T(x_T)\right].$$4 a strength, $$J(\mathbf u)=\mathbb E\!\left[\sum_{t=0}^{T-1} c(x_t,u_t)+c_T(x_T)\right].$$5 a feature type, and $$J(\mathbf u)=\mathbb E\!\left[\sum_{t=0}^{T-1} c(x_t,u_t)+c_T(x_T)\right].$$6 a direction vector used for some types. The feature types are local minima, low-gradient regions, and high-curvature regions. Memory evolves according to

$$J(\mathbf u)=\mathbb E\!\left[\sum_{t=0}^{T-1} c(x_t,u_t)+c_T(x_T)\right].$$7

with $$J(\mathbf u)=\mathbb E\!\left[\sum_{t=0}^{T-1} c(x_t,u_t)+c_T(x_T)\right].$$8 denoting extracted topological information from the current state or trajectory context (Zheng et al., 24 Sep 2025).

The memory potential is a sum of local feature potentials,

$$J(\mathbf u)=\mathbb E\!\left[\sum_{t=0}^{T-1} c(x_t,u_t)+c_T(x_T)\right].$$9

For local minima,

$$u_t^*=\int_{\mathcal T} u_t(\tau)\,p(\tau\mid x_t)\,d\tau, \qquad p(\tau\mid x_t)=\frac{1}{Z(x_t)}\exp\!\left(-\frac{1}{\lambda}S(\tau)\right),$$0

For low-gradient regions,

$$u_t^*=\int_{\mathcal T} u_t(\tau)\,p(\tau\mid x_t)\,d\tau, \qquad p(\tau\mid x_t)=\frac{1}{Z(x_t)}\exp\!\left(-\frac{1}{\lambda}S(\tau)\right),$$1

For high-curvature regions,

$$u_t^*=\int_{\mathcal T} u_t(\tau)\,p(\tau\mid x_t)\,d\tau, \qquad p(\tau\mid x_t)=\frac{1}{Z(x_t)}\exp\!\left(-\frac{1}{\lambda}S(\tau)\right),$$2

These potentials are designed to produce repulsion from traps, directional guidance across plateaus, and saddle-like channels through narrow passages (Zheng et al., 24 Sep 2025).

The blending coefficient is given in one form by

$$u_t^*=\int_{\mathcal T} u_t(\tau)\,p(\tau\mid x_t)\,d\tau, \qquad p(\tau\mid x_t)=\frac{1}{Z(x_t)}\exp\!\left(-\frac{1}{\lambda}S(\tau)\right),$$3

and in another by

$$u_t^*=\int_{\mathcal T} u_t(\tau)\,p(\tau\mid x_t)\,d\tau, \qquad p(\tau\mid x_t)=\frac{1}{Z(x_t)}\exp\!\left(-\frac{1}{\lambda}S(\tau)\right),$$4

The paper presents these as equivalent design variants for smooth switching between base and memory control (Zheng et al., 24 Sep 2025).

3. Memory as a field-shaping mechanism

Across the related literature, the distinctive feature of MAPFT is not merely the presence of memory, but the conversion of memory into a continuous field that can influence control locally.

In the neural field formulation, the position layer $$u_t^*=\int_{\mathcal T} u_t(\tau)\,p(\tau\mid x_t)\,d\tau, \qquad p(\tau\mid x_t)=\frac{1}{Z(x_t)}\exp\!\left(-\frac{1}{\lambda}S(\tau)\right),$$5 encodes the agent’s current position as a localized bump attractor, while the memory layer $$u_t^*=\int_{\mathcal T} u_t(\tau)\,p(\tau\mid x_t)\,d\tau, \qquad p(\tau\mid x_t)=\frac{1}{Z(x_t)}\exp\!\left(-\frac{1}{\lambda}S(\tau)\right),$$6 stores previously visited locations as persistent activity bounded by a front (Kilpatrick et al., 2017). Search is then biased through a closed-loop modulation

$$u_t^*=\int_{\mathcal T} u_t(\tau)\,p(\tau\mid x_t)\,d\tau, \qquad p(\tau\mid x_t)=\frac{1}{Z(x_t)}\exp\!\left(-\frac{1}{\lambda}S(\tau)\right),$$7

so that overlap between position and memory affects the effective velocity input. The reduced interface dynamics show that the bump center obeys

$$u_t^*=\int_{\mathcal T} u_t(\tau)\,p(\tau\mid x_t)\,d\tau, \qquad p(\tau\mid x_t)=\frac{1}{Z(x_t)}\exp\!\left(-\frac{1}{\lambda}S(\tau)\right),$$8

and the front boundaries advance only when the bump is appropriately positioned relative to the memory front. The resulting landscape is history-dependent and state-dependent rather than a fixed scalar potential. The paper explicitly notes that this is analogous to classical potential-field navigation, but differs because the “potential landscape” is generated dynamically by neural memory (Kilpatrick et al., 2017).

In the reinforcement-field formulation, the field is defined in Hilbert space by kernels centered on remembered augmented states. The augmented state space is $$u_t^*=\int_{\mathcal T} u_t(\tau)\,p(\tau\mid x_t)\,d\tau, \qquad p(\tau\mid x_t)=\frac{1}{Z(x_t)}\exp\!\left(-\frac{1}{\lambda}S(\tau)\right),$$9, the fitness function is $S(\tau)$0, and an experience particle is the tuple $S(\tau)$1. The reinforcement field is then “a vector field in Hilbert space established by one or more kernels through their linear combination as a representation for the fitness function,” with Gaussian process prediction

$S(\tau)$2

and kernel expansion

$S(\tau)$3

Positive and negative particle polarity is determined by the sign of the temporal difference, so memory elements behave like supportive or inhibitory sources in the induced field (Chiu et al., 2022).

In APF-augmented LMAPF, the relevant memory is spatial-temporal and collective. For TA$S(\tau)$4+APF, if $S(\tau)$5 is the location of agent $S(\tau)$6 at time $S(\tau)$7, then the repulsive potential around that trajectory is

$S(\tau)$8

and the total APF penalty is

$S(\tau)$9

The paper emphasizes that APFs should be added to the $\lambda>0$0-value rather than the heuristic, because only the accumulated path cost can change which path is optimal under A$\lambda>0$1-style search (Pertzovsky et al., 28 May 2025). A plausible implication is that, in MAPFT terms, memory is operational only when it is embedded into the dynamics of path evaluation, not when it remains a purely advisory signal.

4. Algorithms and controller instantiations

The principal controller instantiation of MAPFT is Memory-Augmented MPPI (MA-MPPI), which contains four modules: MPPI control core, topological feature detector, memory representation, and adaptive potential field synthesizer (Zheng et al., 24 Sep 2025). It uses the memory-augmented value

$\lambda>0$2

and modifies exploration through

$\lambda>0$3

with covariance scaling

$\lambda>0$4

Optimal control is estimated by weighted averaging,

$\lambda>0$5

Thus memory acts in two places simultaneously: it reshapes the cost landscape and it widens exploration near memorized difficult regions (Zheng et al., 24 Sep 2025).

The reinforcement-field literature gives a different algorithmic realization. RF-SARSA combines a SARSA base learner,

$\lambda>0$6

with Gaussian-process fitness prediction over experience particles and softmax action selection,

$\lambda>0$7

MemoryUpdate maintains the working-memory particle set through correlation and polarity checks, while G-SARSA extends the system to abstract actions and “decision concepts” via spectral clustering (Chiu et al., 2022).

The search-theoretic neural field model is analytically reduced to interface equations for bump and front boundaries, allowing the PDE system to be represented by low-dimensional dynamics that track current position and memory-front location (Kilpatrick et al., 2017). In multi-agent planning, APF augmentation is layered into TA$\lambda>0$8, SIPPS, PIBT, and LaCAM rather than used as a standalone planner; direct APF is fast but too myopic in dense maps, whereas APF as a bias on top of a planner that already ensures feasibility or completeness functions as a practical congestion-memory mechanism (Pertzovsky et al., 28 May 2025).

5. Theoretical properties and empirical results

The explicit MAPFT paper states three central theoretical claims. Theorem 1 gives a non-convex escape property: if a local minimum region $\lambda>0$9 is recorded in memory and the memory strength satisfies

$$V(x,M)=\alpha(x,M)\,V_{\text{base}}(x)+(1-\alpha(x,M))\,V_{\text{mem}}(x,M).$$0

then for any $$V(x,M)=\alpha(x,M)\,V_{\text{base}}(x)+(1-\alpha(x,M))\,V_{\text{mem}}(x,M).$$1, there exists finite $$V(x,M)=\alpha(x,M)\,V_{\text{base}}(x)+(1-\alpha(x,M))\,V_{\text{mem}}(x,M).$$2 such that escape from the region occurs with probability at least $$V(x,M)=\alpha(x,M)\,V_{\text{base}}(x)+(1-\alpha(x,M))\,V_{\text{mem}}(x,M).$$3. The proof idea is that the memory gradient creates an outward force that dominates the inward pull of the base objective inside the region, yielding a finite escape time bound of the form

$$V(x,M)=\alpha(x,M)\,V_{\text{base}}(x)+(1-\alpha(x,M))\,V_{\text{mem}}(x,M).$$4

Theorem 2 gives asymptotic convergence under coercivity of $$V(x,M)=\alpha(x,M)\,V_{\text{base}}(x)+(1-\alpha(x,M))\,V_{\text{mem}}(x,M).$$5, and Theorem 3 states that if there are $$V(x,M)=\alpha(x,M)\,V_{\text{base}}(x)+(1-\alpha(x,M))\,V_{\text{mem}}(x,M).$$6 independent local minimum regions, then the expected time for standard MPPI is lower bounded relative to MA-MPPI by

$$V(x,M)=\alpha(x,M)\,V_{\text{base}}(x)+(1-\alpha(x,M))\,V_{\text{mem}}(x,M).$$7

The paper connects these results to Morse theory, topological surgery intuition, and persistent homology (Zheng et al., 24 Sep 2025).

The reported empirical evidence spans robotic control, engineering control, and navigation:

Domain	Reported metrics	Reported outcome
Robotic control benchmarks	cumulative reward, sample efficiency $$V(x,M)=\alpha(x,M)\,V_{\text{base}}(x)+(1-\alpha(x,M))\,V_{\text{mem}}(x,M).$$8, local optima escape rate, trap frequency, computation time, control smoothness	MA-MPPI outperformed all baselines across Pendulum-v1, BipedalWalker-v3, HalfCheetah-v4, and Humanoid-v4
Power system control	constraint violation rate, economic efficiency, stability margin, compute time, recovery time	MA-MPPI reduced violations and nearly halved recovery time on the IEEE 39-bus New England system
UAV obstacle avoidance	success rate, path optimality, control smoothness, compute time, local minima escapes	MA-MPPI was more successful, smoother, and far better at escaping traps

More specifically, MA-MPPI achieved mean rewards of $$V(x,M)=\alpha(x,M)\,V_{\text{base}}(x)+(1-\alpha(x,M))\,V_{\text{mem}}(x,M).$$9 on Pendulum-v1, $V_{\text{base}}(x)$0 on BipedalWalker-v3, $V_{\text{base}}(x)$1 on HalfCheetah-v4, and $V_{\text{base}}(x)$2 on Humanoid-v4, compared with MPPI values of $V_{\text{base}}(x)$3, $V_{\text{base}}(x)$4, $V_{\text{base}}(x)$5, and $V_{\text{base}}(x)$6, respectively. Sample efficiency to reach $V_{\text{base}}(x)$7 asymptotic performance was reported as 78, 183, 324, and 568 interactions across the same tasks, compared with MPPI values of 124, 352, 586, and 984 and SAC values of 267, 624, 1248, and 1875. Escape rates from trap states ranged from $V_{\text{base}}(x)$8 on Pendulum-v1 to $V_{\text{base}}(x)$9 on Humanoid-v4, versus $V_{\text{mem}}(x,M)$0 to $V_{\text{mem}}(x,M)$1 for standard MPPI. Runtime overhead was about $V_{\text{mem}}(x,M)$2 relative to MPPI (Zheng et al., 24 Sep 2025).

In the IEEE 39-bus benchmark, MA-MPPI versus MPPI yielded constraint violations of $V_{\text{mem}}(x,M)$3 versus $V_{\text{mem}}(x,M)$4, economic efficiency $V_{\text{mem}}(x,M)$5 versus $V_{\text{mem}}(x,M)$6, stability margin $V_{\text{mem}}(x,M)$7 versus $V_{\text{mem}}(x,M)$8, compute time $V_{\text{mem}}(x,M)$9 versus $\alpha(x,M)\in[0,1]$0, and disturbance recovery $\alpha(x,M)\in[0,1]$1 versus $\alpha(x,M)\in[0,1]$2. In UAV obstacle avoidance, success rate was $\alpha(x,M)\in[0,1]$3 versus $\alpha(x,M)\in[0,1]$4, path optimality $\alpha(x,M)\in[0,1]$5 versus $\alpha(x,M)\in[0,1]$6, control smoothness $\alpha(x,M)\in[0,1]$7 versus $\alpha(x,M)\in[0,1]$8, compute time $\alpha(x,M)\in[0,1]$9 versus $\alpha\approx 1$0, and local minima escapes $\alpha\approx 1$1 versus $\alpha\approx 1$2 (Zheng et al., 24 Sep 2025).

The related literatures show a more qualified picture. In the neural search model, memory-guided speed changes do not improve performance on a single segment; the optimal strategy is effectively $\alpha\approx 1$3. By contrast, inhibition-of-return is provably better than random arm choice in a radial arm maze, with

$\alpha\approx 1$4

Thus memory helps in multi-branch or multi-arm environments, but not necessarily on a single 1D segment (Kilpatrick et al., 2017). Likewise, in multi-agent planning, APFs are not beneficial for one-shot MAPF, but in LMAPF they yield up to a 7-fold increase in overall system throughput; on the room map, PIBT+APF improves throughput by around $\alpha\approx 1$5 on average over PIBT, and APF-augmented methods are especially effective under RHCR because repeated short-horizon replanning makes congestion memory actionable (Pertzovsky et al., 28 May 2025).

6. Distinctions, limitations, and open issues

A persistent misconception is that MAPFT is simply classical artificial potential fields with a memory buffer. The literature is more specific. Classical APFs usually impose a handcrafted scalar potential $\alpha\approx 1$6 and motion follows a gradient flow, whereas the neural field model generates its effective landscape from persistent neural memory (Kilpatrick et al., 2017); the reinforcement-field model learns a Bayesian, kernel-induced interpolation surface over remembered particles rather than a hand-designed geometric potential (Chiu et al., 2022); and the explicit MAPFT framework learns feature instances from trajectory history and uses them to modify both cost and sampling in stochastic optimal control (Zheng et al., 24 Sep 2025).

A second misconception is that memory augmentation is uniformly beneficial. The reported results are domain-dependent. Memory-guided speed modulation does not improve the single-segment search task (Kilpatrick et al., 2017). APFs are “not very helpful for one-shot, offline MAPF” and can even worsen sum-of-costs, although they substantially improve throughput in LMAPF (Pertzovsky et al., 28 May 2025). This suggests that memory fields are most useful when the task is recurrent, non-convex, congested, or history-sensitive, and less useful when a single feasible solution in a static environment is the primary objective.

The current limitations are also explicit. The MAPFT paper notes limited generalization between similar features, no sophisticated long-horizon memory management, and no multi-agent knowledge sharing; the appendix argues that better generalization may require hierarchical feature representations, manifold-aware similarity, learned embeddings, and transfer learning across features (Zheng et al., 24 Sep 2025). The reinforcement-field framework acknowledges that memory updates and kernel inversion can be computationally expensive, that cluster reassignment and concept drift are open issues, and that kernel choice, thresholds, partitioning, and update periods remain heuristic (Chiu et al., 2022). The neural field model is analytically strongest in one dimension, with higher-dimensional analysis mostly proposed rather than fully derived, and its use of Heaviside nonlinearities and idealized kernels is biologically coarse (Kilpatrick et al., 2017). In APF-augmented LMAPF, direct APF can get trapped in local minima or corridor deadlocks, while APF integration introduces overhead of roughly $\alpha\approx 1$7 for TA$\alpha\approx 1$8+APF, $\alpha\approx 1$9 for SIPPS+APF, and $$J(\mathbf u)=\mathbb E\!\left[\sum_{t=0}^{T-1} c(x_t,u_t)+c_T(x_T)\right].$$00 for PIBT+APF, with substantial sensitivity to $$J(\mathbf u)=\mathbb E\!\left[\sum_{t=0}^{T-1} c(x_t,u_t)+c_T(x_T)\right].$$01, $$J(\mathbf u)=\mathbb E\!\left[\sum_{t=0}^{T-1} c(x_t,u_t)+c_T(x_T)\right].$$02, $$J(\mathbf u)=\mathbb E\!\left[\sum_{t=0}^{T-1} c(x_t,u_t)+c_T(x_T)\right].$$03, and $$J(\mathbf u)=\mathbb E\!\left[\sum_{t=0}^{T-1} c(x_t,u_t)+c_T(x_T)\right].$$04 (Pertzovsky et al., 28 May 2025).

Taken together, these results position MAPFT as a general theory of history-dependent landscape shaping. In one formulation, memory is a set of topological feature instances; in another, a pinned front over visited space; in another, a kernelized cloud of experience particles; and in another, a repulsive trace of recent traffic. Across these formulations, the common principle is that past interaction is converted into a structured field that biases future control, search, or planning in a way that standard memoryless methods do not capture (Zheng et al., 24 Sep 2025).