Memory as Structured Trajectories: Persistent Homology and Contextual Sheaves
Xin Li, Department of Computer ScienceUniversity at Albany [email protected] · 2025
Read Memory as Structured Trajectories: Persistent Homology and Contextual Sheaves by Xin Li, Department of Computer ScienceUniversity at Albany [email protected] (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-11646.txt.
Introduction
Xin Li asks a clean question: what should a memory system keep, and what should it drop? Nerve and immune memory face the same bind: the future rarely repeats the past word for word. Keeping every detail wastes room, while dropping the wrong detail costs the right response later.
Li frames memory as scaffold and flow: fast responses flow, while slow body traits form a scaffold that shapes them. Consolidation writes a coarse summary of the past into that scaffold. A good summary keeps splits that change future acts, stays cheap, and runs on its own. A leakage score counts where coarse dynamics fail. Across three test systems, from nerve nets to immune tuning, mid-level grain wins. Too fine wastes space, too coarse merges cases that need distinct replies.
Catalog record
| CREATOR | Xin Li, Department of Computer ScienceUniversity at Albany [email protected] |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 109 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-11646
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Memory as Structured Trajectories: Persistent Homology and Contextual Sheaves Xin Li, Department of Computer ScienceUniversity at Albany [email protected] Abstract We propose a topological framework for memory and inference grounded in the structure of spike-timing dynamics, persistent homology, and the Context-Content Uncertainty Principle (CCUP). Starting from the observation that polychronous neural groups (PNGs) encode reproducible, time-locked spike sequences shaped by axonal delays and synaptic plasticity, we construct spatiotemporal complexes whose temporally consistent transitions define chain complexes over which robust activation cycles emerge. These activation loops are abstracted into cell posets , enabling a compact and causally ordered representation of neural activity with overlapping and compositional memory traces. We introduce the delta-homology analogy , which formalizes memory as a set of sparse, topologically irreducible attractors. A Dirac delta-like memory trace δ γ \delta_{\gamma} is identified with a nontrivial homology generator [ γ ] ∈ H 1 ( 𝒵 ) [\gamma]\in H_{1}(\mathcal{Z}) on a latent manifold of cognitive states. Such traces are sharply localized along reproducible topological cycles and are only activated when inference trajectories complete a full cycle. They encode minimal, path-dependent memory units that cannot be synthesized from local features alone. We interpret these delta-homology generators as the low-entropy content variable Φ \Phi , while the high-entropy context variable Ψ \Psi is represented dually as a filtration, cohomology class, or sheaf over the same latent space. Inference is recast as a dynamic alignment between Φ \Phi and Ψ \Psi , and coherent memory retrieval corresponds to the existence of a global section that selects and sustains a topological generator. When contextual coherence fails, inference becomes fragmented and non-integrable, giving rise to ambiguity, hallucination, or misbinding. This framework unifies spiking dynamics, topological persistence, and contextual modulation under the CCUP. It leads to a new model of memory in which inference, learning, and recall arise from structured, reproducible trajectories through a delta-homology space, constrained by the geometry of uncertainty and modulated by context. Memory is no longer a static attractor or distributed code, but a cycle-completing, structure-aware inference process. Keywords: delta-homology analogy, polychronous neural group (PNG), cell poset, persistent homology, homology generator, sheaf coherence, memory-amortized inference (MAI), homology-cohomology loop I Introduction How does the brain represent and generalize memory traces that are context-sensitive and robust to perturbation? Classical models of memory, ranging from attractor networks (e.g., Hopfield network [ 1 , 2 ] ) to distributed embeddings (e.g., key-value networks [ 3 , 4 ] ), tend to emphasize static fixed points or continuous similarity-based interpolation in high-dimensional spaces. However, mounting evidence suggests that biological memory is inherently dynamic, path-dependent, and temporally compositional [ 5 , 6 ] . In particular, polychronous neural groups (PNGs) [ 7 ] , reproducible spike sequences shaped by fixed axonal delays and spike-timing-dependent plasticity (STDP) [ 8 ] , offer a biologically grounded model of structured memory that unfolds over time, rather than being localized in space. These polychronous activation patterns trace closed, temporally precise loops through latent sensorimotor spaces [ 9 ] . Unlike synchronous assemblies, PNGs are irreducible to static co-activation: they encode reproducible causal chains of spike-timing relationships that persist across trials and contexts [ 10 ] . To capture this structure, we construct spatiotemporal complexes whose edges encode delay-respecting transitions, and then extract topological features through chain complexes and homology [ 11 ] . The partially ordered and overlapping spike chains are then abstracted into cell posets [ 12 ] , generating a combinatorial scaffold on which topological memory can be built. Chain complexes over posets can be used to compute persistent topological features [ 13 ] ; and sheaves or filtrations can be defined over posets to modulate activation [ 14 ] . Cycles as Fundamental Memory Units. We then introduce the concept of delta-homology , a filtration-based generalization of classical homology that measures the persistence of topological features across spike-timing tolerance (e.g., the filtration parameter). Within this framework, we define the delta-homology analogy : each minimal, sharply localized memory trace is represented by a Dirac delta-like distribution \delta_{\gamma} [ 15 ] that corresponds to a nontrivial homology generator [\gamma]\in H_{1}(\mathcal{Z}) , where \mathcal{Z} is a latent manifold of cognitive states [ 14 ] . Such traces activate only when the inference trajectory completes the full cycle \gamma ; they cannot be synthesized from local cues alone and represent sparse, non-interpolable, path-dependent memory units. A Dirac delta-like memory trace \delta_{\gamma} corresponds to a nontrivial topological generator [\gamma]\in H_{1}(\mathcal{Z}) , and becomes active only when the inference trajectory completes a globally coherent loop. From the perspective of sheaf theory [ 16 ] , local spike-timing constraints define sections over patches of a cell poset on a latent manifold. The memory trace is activated only when these local sections glue into a consistent global section - i.e., when local inferences align to form a non-contractible cycle. Such traces are non-interpolable : no subset of local activations can synthesize them [ 17 ] . Their emergence requires both structural closure and contextual coherence. Failure to meet this condition manifests as cohomological obstruction H^{1}(\mathcal{F})\neq 0 , fragmentation, or hallucinated inference. To embed this structural substrate into a functional model of memory, we invoke the Context-Content Uncertainty Principle (CCUP) [ 18 ] . CCUP posits a dual decomposition of latent variables into a low-entropy content variable \Phi , represented by persistent delta-homology generators encoding structured memory; and a high-entropy context variable \Psi , modeled as a filtration, cohomology class, or sheaf that dynamically modulates which content cycles are activated or suppressed [ 16 ] . Under this framework, inference is framed as a cycle of mutual alignment between context and content: \Phi provides reproducible latent structure, while \Psi defines the conditions under which that structure becomes relevant. Coherent inference corresponds to the existence of a global section in the contextual sheaf that supports the activation of a topological memory generator [ 19 ] . An incoherent or fragmented context leads to disorganized or ambiguous retrieval, reflecting failures of inference integration. Topology Before Geometry. The use of homology as a mathematical language for describing PNGs aligns with Jean Piaget’s developmental theory of spatial cognition [ 20 ] . In his seminal work on the child’s conception of space, Piaget observed that children acquire topological concepts, such as connectedness, enclosure, and continuity, prior to understanding metric geometric notions like distance, angle, and coordinate systems [ 21 ] . This topological primacy reflects a developmental ordering in which qualitative, structure-preserving relationships form the foundation for later quantitative refinements. Within the delta-homology framework, memory traces emerge as persistent 1-cycles in a filtered spatiotemporal complex, capturing reproducible, temporally constrained patterns of neural activity. These cycles, corresponding to homology generators, are topological invariants that encode path-dependent structure independently of any metric embedding. The topology-first architecture of PNG-based memory resonates with Piaget’s insight [ 22 ] : cognition is grounded in the preservation of structure before the imposition of metric form, and meaningful memory arises from the recognition of loops, boundaries, and connectivity, core topological constructs that precede geometric inference. The rest of this paper is organized as follows. Section II rigorously formalizes the temporal structure of PNG into spatiotemporal complex as a stepping stone for bridging chain complex in algebraic topology. Section