Sujal Srivastava · Aug 15, 2026
Modern enterprise information systems contain heterogeneous streams of information that differ
fundamentally in representation, temporal structure, and semantics. Relational databases encode
structured attributes and transactions; knowledge graphs encode entities and relations; textual
systems encode sequential symbolic information; telemetry and event streams additionally encode
temporal dynamics. Conventional memory systems typically transform these modalities into a
common vector space and subsequently perform similarity search, graph traversal, or learned re-
trieval. Such approaches provide effective retrieval but do not naturally model one-shot associative
learning, temporal binding, pattern completion, continual online learning, or biologically inspired
memory consolidation.
This work proposes a Polychronous Multimodal Episodic Memory (PMEM) architecture based
on a spiking neural substrate inspired by the entorhinal cortex (EC), dentate gyrus (DG), hip-
pocampal CA3, and CA1. Heterogeneous modalities are independently transformed into tempo-
rally structured spike trains and converged into a common entorhinal representation. The DG
performs sparse pattern separation, while recurrent CA3 dynamics form autoassociative attractor
states capable of completing memories from partial or cross-modal cues. Spike-timing-dependent
plasticity (STDP), neuromodulatory gating, synaptic delays, and polychronous convergence pro-
vide mechanisms for rapid associative binding. CA1 provides a learned readout from the latent
hippocampal state back to modality-specific representations. Subsequent extensions introduce
episodic temporal sequences, replay-based consolidation, continual learning, homeostatic plastic-
ity, modular scaling, and cross-modal retrieval.
The resulting system is not intended to replace relational databases, graph databases, or lan-
guage models. Instead, it defines a neural associative memory substrate operating over these
systems, in which structured records, graph relations, and textual episodes become temporally en-
coded neural events. The central research hypothesis is that multimodal information can be bound
through spike timing and recurrent dynamics without requiring a single shared semantic embed-
ding or an explicit retrieval index. The current prototype demonstrates encoder reconstruction,
CA3 pattern completion, multimodal binding, engram stability, novelty-sensitive plasticity, and
continuous two-stage encoding; however, several components remain prototype-level, particularly
pure neural retrieval without symbolic assistance, large-scale capacity, semantic text encoding,
continual consolidation, and rigorous comparison against modern memory baselines
COMPUTER SCIENCE AND ENGINEERING