Topological Boundary Memory

• Topological boundary memory is a class of mechanisms where stable, nonlocal boundary states store and retrieve information based on topological invariants.
• It leverages diverse platforms—magnetic, quantum, photonic, and mechanical—to realize robust memory registers that resist disorder and imperfections.
• Research shows that manipulating boundary conditions via quenches, interfaces, and protocol changes yields measurable, device-level memory effects.

Searching arXiv for relevant papers on topological boundary memory and related boundary-based/topological memory mechanisms. Topological boundary memory denotes a family of memory mechanisms in which stored, retrievable, or protected information is tied to a boundary object—such as a surface state, edge mode, domain wall, synthetic edge, or boundary algebra—whose stability is governed by topology rather than by a purely local bulk variable. In the literature surveyed here, the phrase encompasses magnetically gapped topological-insulator surfaces read out through a half-quantized Hall response, Majorana edge modes used as a quantum memory register, boundary-condition-driven edge occupations in one-dimensional topological chains, boundary-engineered defect strings in frustrated media, and operator-algebraic boundary nets whose superselection sectors recover topological boundary order (Fujita et al., 2011, Bedow et al., 13 May 2025, He et al., 2016, Jones et al., 24 Jun 2025).

1. Conceptual range of the boundary

Across these works, “boundary” is not restricted to a literal sample edge. It can mean a physical surface of a 3D topological insulator, an interface between topological and trivial domains in a superconductor or photonic crystal, a domain wall between competing ordered regions, a synthetic edge in Floquet photon space, a cycle that has vanishing boundary but is not itself a boundary in homology, or a boundary net of local algebras extracted from a bulk quantum spin system (Fujita et al., 2011, Baum et al., 2017, Li, 1 Aug 2025, Jones et al., 24 Jun 2025).

A recurring structural distinction is between trivial closure and nontrivial closure. In the neural-homological formulation, a memory trace is required to satisfy

$$\gamma \in \ker \partial_1 \quad \text{and} \quad \gamma \notin \operatorname{im}\partial_2,$$

so it is closed but non-bounding (Li, 1 Aug 2025). In topological mechanics, a mismatch of bulk winding data across an edge or domain wall forces boundary-localized floppy modes or self-stress states (Kane et al., 2013). In electronic and photonic settings, bulk-boundary correspondence ties quantized bulk invariants to boundary conductance or edge transport (Fujita et al., 2011, Ahmadnejad et al., 26 Feb 2025).

Setting	Boundary object	Memory-bearing quantity
Magnetic topological insulator	Gapped TI surface	Sign of surface magnetization via Hall response
Majorana platform	Domain boundary / edge mode	MEM occupation transferred to MZMs
Photonic Chern system	Edge channel at interface	Chern phase and edge transport
Colloidal ice	Defect string between domains	Boundary-imprinted string configuration
Interacting SSH/XXZ chain	First dimer / edge magnetization	Post-quench boundary polarization
Algebraic bulk-boundary order	Boundary net of algebras	DHR sector data

This diversity also constrains the meaning of the term. The literature does not present a single universal formalism; rather, it presents a common pattern in which topology stabilizes information at an interface, and the interface provides either the memory register, the readout channel, or the reconstruction of the relevant topological data.

2. Electronic and quantum-device realizations

A canonical device-level realization is the topological-insulator memory cell proposed on a magnetically doped surface of a 3D topological insulator (Fujita et al., 2011). The surface is described by

$$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$

with spectrum

$$E_\tau=\tau\sqrt{v_F^2 p^2+m^2}, \qquad \tau=\pm 1,$$

and gap

$\Delta = 2|m|.$

The bit is encoded by the sign of the perpendicular magnetization, for example upward magnetization as “1” and downward magnetization as “0”. When the Fermi level lies in the gap, the Hall conductivity takes the half-quantized form

$\sigma_{xy}=-\frac{e^2}{2h}\,\text{sgn}(m),$

which the paper explicitly identifies as a topological invariant (Fujita et al., 2011). Writing is conventional, by applying a magnetic field larger than the coercive field; the cited coercivity for magnetically doped Bi$_2$Te$_3$ is $H_C \sim 0.01\ \text{T}$. Readout is electrical through the Hall voltage, with an estimate of $V_H \approx \pm 47~\text{mV}$ for $I=1~\mu\text{A}$ at low temperature. The proposal emphasizes robustness against disorder and impurities, device-geometry variations, and imperfections in writing, precisely because the readout is tied to a topological invariant rather than to a thresholded comparator signal (Fujita et al., 2011).

The same paper makes the operating limits explicit. Magnetic impurity doping such as Mn-doped Bi$$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$0Te$$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$1 can produce a gap of order $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$2 meV, while a ferromagnetic overlayer typically produces a smaller gap around $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$3 meV, albeit with potentially higher Curie temperature. Quantized readout requires $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$4, and the discussion notes experimentally observed gaps around $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$5 meV in Mn-doped Bi$$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$6Se$$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$7, corresponding to roughly $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$8 K, while also noting reported Curie temperatures below about $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$9 K in some systems (Fujita et al., 2011). A common misconception is that topology by itself guarantees unrestricted operating temperature; this proposal states the opposite, namely that quantization precision degrades once thermal excitations smear the occupations.

A distinct quantum-information realization appears in the use of Majorana edge modes as quantum memory in a two-dimensional topological superconductor (Bedow et al., 13 May 2025). In that architecture, topological and trivial domains coexist; Majorana edge modes (MEMs) appear on domain boundaries, while Majorana zero modes (MZMs) appear in vortex cores once vortices enter the topological region. The memory mechanism is explicit: when vortices are in the trivial region, the qubit is stored in the occupation of delocalized MEMs along the domain boundary; when vortices cross into the topological region, the MEMs are converted into localized vortex-core MZMs; after the gate operation, the vortices return and the state is again encoded in the MEMs (Bedow et al., 13 May 2025). The adiabaticity requirement is written as

$$E_\tau=\tau\sqrt{v_F^2 p^2+m^2}, \qquad \tau=\pm 1,$$0

to suppress leakage into Caroli–de Gennes–Matricon states.

The same work ties boundary memory to gate synthesis. It reports $$E_\tau=\tau\sqrt{v_F^2 p^2+m^2}, \qquad \tau=\pm 1,$$1- and $$E_\tau=\tau\sqrt{v_F^2 p^2+m^2}, \qquad \tau=\pm 1,$$2-gate implementations with geometric phase differences $$E_\tau=\tau\sqrt{v_F^2 p^2+m^2}, \qquad \tau=\pm 1,$$3 and $$E_\tau=\tau\sqrt{v_F^2 p^2+m^2}, \qquad \tau=\pm 1,$$4, respectively, and an $$E_\tau=\tau\sqrt{v_F^2 p^2+m^2}, \qquad \tau=\pm 1,$$5-gate for which

$$E_\tau=\tau\sqrt{v_F^2 p^2+m^2}, \qquad \tau=\pm 1,$$6

For an $$E_\tau=\tau\sqrt{v_F^2 p^2+m^2}, \qquad \tau=\pm 1,$$7 extension, the state evolves into an equal superposition of the 8 even-parity states with

$$E_\tau=\tau\sqrt{v_F^2 p^2+m^2}, \qquad \tau=\pm 1,$$8

The paper’s interpretation is that MEMs are not merely passive edge excitations; they are “functionalized as quantum memory” (Bedow et al., 13 May 2025).

3. Boundary changes, quenches, and protocol memory

In one-dimensional topological systems, memory can be carried not by a static boundary register but by the system’s retained dependence on how a boundary was changed. For the Kitaev chain and SSH model, changing only one boundary link from periodic to open induces a disturbance that propagates into the bulk with a light-cone structure, while the density of topological edge modes reaches a steady state whose value depends on the boundary ramp time $$E_\tau=\tau\sqrt{v_F^2 p^2+m^2}, \qquad \tau=\pm 1,$$9 (He et al., 2016). In the Kitaev chain, the boundary protocol is

$\Delta = 2|m|.$0

followed by $\Delta = 2|m|.$1 for $\Delta = 2|m|.$2. The topological criterion is $\Delta = 2|m|.$3. The final edge-mode density $\Delta = 2|m|.$4 approaches $\Delta = 2|m|.$5 as $\Delta = 2|m|.$6 increases, and $\Delta = 2|m|.$7 is approximately linear in $\Delta = 2|m|.$8 (He et al., 2016).

This paper gives a second memory signature: some correlations remain finite across the link that has been physically broken, but only in the topological regime. In the Kitaev chain, $\Delta = 2|m|.$9 decays but remains finite after the link is broken, whereas it decays to zero in the trivial regime. In the SSH model, the analogous quantity is $\sigma_{xy}=-\frac{e^2}{2h}\,\text{sgn}(m),$0, which likewise remains finite only when edge modes exist (He et al., 2016). The result is explicitly contrasted with non-topological regimes, where none of these boundary-memory signatures survive.

A stronger nonequilibrium result appears in the bond-alternating XXZ chain, an interacting SSH model hosting symmetry-protected topological edge modes (Hang et al., 17 Jun 2026). The central organizing principle is whether the post-quench Hamiltonian is free or genuinely interacting. For a free post-quench Hamiltonian, the left boundary-mode return amplitude satisfies

$\sigma_{xy}=-\frac{e^2}{2h}\,\text{sgn}(m),$1

so the boundary memory decays algebraically (Hang et al., 17 Jun 2026). For a genuinely interacting post-quench Hamiltonian, the paper proves finite-time stability bounds. The first-dimer magnetization obeys

$\sigma_{xy}=-\frac{e^2}{2h}\,\text{sgn}(m),$2

and, away from local resonances, one can construct a dressed boundary charge $\sigma_{xy}=-\frac{e^2}{2h}\,\text{sgn}(m),$3 such that the memory remains stable on time windows growing as arbitrarily large powers of the inverse inter-dimer coupling (Hang et al., 17 Jun 2026). The analysis also identifies a local suppression near the isotropic $\sigma_{xy}=-\frac{e^2}{2h}\,\text{sgn}(m),$4 point $\sigma_{xy}=-\frac{e^2}{2h}\,\text{sgn}(m),$5, where a resonant leakage channel opens.

Memory can also generate the boundary itself. In Floquet-induced synthetic crystals, a non-Markovian periodically driven Schrödinger equation,

$\sigma_{xy}=-\frac{e^2}{2h}\,\text{sgn}(m),$6

maps to a synthetic lattice in Floquet harmonic index $\sigma_{xy}=-\frac{e^2}{2h}\,\text{sgn}(m),$7, and the memory term produces an $\sigma_{xy}=-\frac{e^2}{2h}\,\text{sgn}(m),$8-dependent potential $\sigma_{xy}=-\frac{e^2}{2h}\,\text{sgn}(m),$9 (Baum et al., 2017). By choosing

$_2$0

the synthetic electric field is screened in a finite region and a mass domain wall is created at $_2$1. The resulting synthetic edge hosts either Kitaev-like zero-energy boundary states in $_2$2 dimensions or chiral Chern-insulator edge states with

$_2$3

in $_2$4 dimensions (Baum et al., 2017). Here the “memory component” of the dynamics is not a stored bit in the usual sense; it is the mechanism that creates the effective boundary supporting the topological state.

4. Engineered interfaces and transport memories

Boundary engineering in frustrated matter provides a classical route to topological boundary memory. In artificial colloidal ice, an antiferromagnetic frontier forces the system rapidly toward the ground state, unlike open or periodic boundaries, while antiferromagnetic domain-wall boundaries with corner defects create bistable topological strings spanning the bulk (Rodríguez-Gallo et al., 2021). The topological charge is

$_2$5

so ice-rule vertices have $_2$6. The especially notable case is the configuration with two opposite defected corners: the system forms two incompatible type-III ground-state regions separated by a diagonal string of type-IV vertices. The paper identifies this diagonal as the topological boundary between the two ground-state domains and describes the resulting string as a topological memory trace of how the system was prepared (Rodríguez-Gallo et al., 2021).

A photonic realization stores information in bulk topology and reads it out through boundary transport rather than optical trapping. In a honeycomb-lattice photonic crystal, the bulk is described by a two-band model and the edge-state count follows

$_2$7

with the Chern number obtained from Berry curvature over the Brillouin zone (Ahmadnejad et al., 26 Feb 2025). Writing is performed by sweeping a synthetic time-dependent magnetic field, for example

$_2$8

across a critical value. The computational study reports GHz-range write speeds of approximately $_2$9–$_3$0 GHz, $_3$1, and a read time

$_3$2

for $_3$3 and $_3$4 m/s (Ahmadnejad et al., 26 Feb 2025). It also reports $_3$5, retention $_3$6, write energy about $_3$7 pJ/bit, read energy about $_3$8 fJ/bit, defect tolerance about $_3$9, and robustness up to about $H_C \sim 0.01\ \text{T}$0 variation in waveguide parameters (Ahmadnejad et al., 26 Feb 2025). The paper’s key distinction is that the memory is “nontrapping”: the stored state is the topological phase itself, not a confined optical pulse.

Transport rather than static storage is the central feature of the topological boundary ratchet realized in an elastic metamaterial (Omidvar et al., 1 Sep 2025). Information is encoded in buckling domains and their domain walls. Neighboring domains act as distinct topological pumps for their Bogoliubov excitations, and the domain wall hosts a boundary-localized mode. Under cyclic loading of the form $H_C \sim 0.01\ \text{T}$1, the boundary mode softens to zero frequency and then becomes imaginary; the ensuing instability moves the wall by two sites, equivalent to advancing the pump phase by $H_C \sim 0.01\ \text{T}$2 (Omidvar et al., 1 Sep 2025). The motion is thus instability-driven and quantized, rather than adiabatic transport of a linear eigenstate.

A related racetrack-memory approach uses magnetic topological insulators as programmable pinning sites in a ferromagnetic nanowire (Hoque et al., 11 Oct 2025). The interfacial pinning arises from

$H_C \sim 0.01\ \text{T}$3

with current-induced torque generated by spin-momentum-locked MTI surface states. In micromagnetic simulations, the critical current for domain-wall shift increases almost linearly with MTI nanobar width, from $H_C \sim 0.01\ \text{T}$4 A/cm$H_C \sim 0.01\ \text{T}$5 for a $H_C \sim 0.01\ \text{T}$6 nm bar to $H_C \sim 0.01\ \text{T}$7 A/cm$H_C \sim 0.01\ \text{T}$8 for an $H_C \sim 0.01\ \text{T}$9 nm bar (Hoque et al., 11 Oct 2025). The paper also reports tunable pinning, exchange-interaction-based pinning, transverse walls, shift current $V_H \approx \pm 47~\text{mV}$0 A/m$V_H \approx \pm 47~\text{mV}$1, and domain-wall speed of about $V_H \approx \pm 47~\text{mV}$2 m/s (Hoque et al., 11 Oct 2025). In this formulation, the boundary is a reconfigurable magnetic-topological interface rather than a fixed lithographic notch.

5. Homological and neural formulations

In neural-topological work, memory is formulated not as a static stored state but as a structured trajectory whose topological nontriviality determines whether recall occurs. Starting from polychronous neural groups, a spatiotemporal complex $V_H \approx \pm 47~\text{mV}$3 is built using the temporal-consistency criterion

$V_H \approx \pm 47~\text{mV}$4

and then converted into a chain complex $V_H \approx \pm 47~\text{mV}$5 with

$V_H \approx \pm 47~\text{mV}$6

(Li, 1 Aug 2025). A persistent memory trace is therefore a 1-cycle that does not bound. The paper’s delta-homology analogy writes a memory trace $V_H \approx \pm 47~\text{mV}$7 as a sharply localized generator supported on a loop $V_H \approx \pm 47~\text{mV}$8, corresponding to a nontrivial class $V_H \approx \pm 47~\text{mV}$9 (Li, 1 Aug 2025).

This framework interprets topological boundary memory in a particularly literal homological sense. Memory is “boundary-sensitive” because it requires the vanishing of the boundary operator $I=1~\mu\text{A}$0, yet is preserved only when it is not itself in $I=1~\mu\text{A}$1. Context-content duality is expressed as

$I=1~\mu\text{A}$2

with a sheaf $I=1~\mu\text{A}$3 over a cell poset $I=1~\mu\text{A}$4. Retrieval corresponds to a global section,

$I=1~\mu\text{A}$5

and failure of gluing is measured by

$I=1~\mu\text{A}$6

which the paper interprets as fragmentation, ambiguity, or hallucinated inference (Li, 1 Aug 2025). The phrase “topological boundary memory” here refers to a memory trace that closes without collapsing.

A related but distinct neural result concerns forgetting and topological recovery from place-cell data (Chowdhury et al., 2017). A dynamic simplicial complex $I=1~\mu\text{A}$7 is built from place-cell assemblies defined by

$I=1~\mu\text{A}$8

and analyzed with zigzag persistence (Chowdhury et al., 2017). Across $I=1~\mu\text{A}$9 dynamic complexes, the error rate for recovering the topology of arenas with $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$00 obstacles is non-monotone in the memory window $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$01, reaching a minimum around $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$02 (Chowdhury et al., 2017). The result is that too much persistence preserves spurious coactivity and false loops, while too little persistence erases genuine obstacle-related loops. Although this work is not boundary-centered in the same geometric sense as surface-state devices, it is directly relevant to the homological meaning of memory persistence and topological irreducibility.

6. Boundary order, rigidity, and self-correction

Topological mechanics supplies an early formal boundary paradigm. In isostatic lattices, the Calladine relation

$$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$03

governs the mismatch between zero modes and states of self-stress, while the equilibrium matrix $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$04 yields a chiral Hamiltonian

$$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$05

placing the system in symmetry class BDI (Kane et al., 2013). In one dimension, the invariant is the winding of $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$06,

$$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$07

For a subsystem $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$08, the zero-mode count decomposes as

$$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$09

where the topological term is controlled by the winding vector $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$10 (Kane et al., 2013). The paper explicitly interprets the boundary modes as a form of topological memory: the bulk lattice “remembers” its topological class through $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$11, and a cut or domain wall reveals that memory as localized floppy modes or self-stress states.

In quantum memories proper, higher-form symmetry can protect a boundary logical qubit. A three-dimensional construction replaces explicit bulk 1-form symmetry enforcement with emergent bulk symmetry from topological order, leaving only boundary symmetry to be enforced (Stahl, 2022). The improvement is scaling: enforcing the relevant symmetry on the boundary requires $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$12 terms instead of $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$13 terms in the bulk (Stahl, 2022). The model combines two 3D toric codes with a 2D toric code on the boundary and imposes boundary generators

$$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$14

As a result, the only allowed boundary logical operations are composite operators,

$$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$15

which are linearly confined (Stahl, 2022). The paper interprets the boundary as a symmetry-protected topological defect.

An operator-algebraic formulation makes the “memory” content of the boundary mathematically precise. For bulk-boundary local topological order, the boundary algebra is organized as a net $$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$16 on a codimension-one cut, and its DHR bimodule category recovers the topological boundary order (Jones et al., 24 Jun 2025). For Levin–Wen boundaries,

$$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$17

while for Walker–Wang boundaries,

$$\mathcal{H}=v_F\vec{\sigma}\cdot(\vec{p}\times\hat{z})+m\sigma^z,$$18

(Jones et al., 24 Jun 2025). The paper also highlights a structural contrast: in the canonical state of the braided categorical net associated with Walker–Wang boundaries, the cone von Neumann algebras are type I with finite-dimensional centers, unlike the type II and III cone algebras found earlier for Levin–Wen bulk models (Jones et al., 24 Jun 2025). In this setting, topological boundary memory is not a device register but a holographic statement: the boundary algebra retains enough information to reconstruct the relevant categorical order.

Taken together, these lines of work show that topology does not confer a single kind of memory. It can stabilize a Hall readout on a gapped surface, preserve protocol dependence in edge occupations and broken-link correlations, sustain domain-wall strings or ratcheted domain transport, encode irreducible cycle completion in homology, linearly confine logical operators on a protected boundary, or permit recovery of boundary order from a net of algebras. The common feature is that information is stored, read, or reconstructed through a boundary structure whose persistence is constrained by nontrivial topology, but the lifetime, operating regime, and exact meaning of “memory” remain platform-dependent (Fujita et al., 2011, He et al., 2016, Li, 1 Aug 2025, Stahl, 2022).