Home / Products / Tenor 4D-EEKG (SYS-02)
RELEASED & OPEN SOURCE 4D-EEKG v2 ARCHITECTURE SYS-02 // CORE MEMORY SUBSTRATE

Tenor: 4D Spatio-Temporal Episodic-Entity Memory

A mathematically grounded cognitive memory substrate for autonomous agents. Replaces brute-force 1M-token KV-cache bloat and vector search amnesia with four-tier continuous temporal manifolds, signed Hebbian plasticity, and deterministic sub-350-token prompt synthesis.

PYTHON 3.10+ SPARSE CSR COMPRESSION PEER-REVIEW PREPRINT
exorticlabs/tenor ★ / 4D-EEKG Foundations Paper ↗
CORE COGNITIVE SUBSTRATE // SYS-02

The Physical Limit of Context Windows

Every autonomous agent running today hits a physical wall: AI amnesia. Massive 1M-token context windows consume over 130 GB of VRAM per session in KV-caches, while vector databases blurt out outdated chunks because semantic similarity has no concept of chronology or state invalidation.

Tenor (4D-EEKG) replaces brute-force prompt stuffing with an integrated four-tier cognitive substrate: immutable temporal intervals [t_start, t_end], signed Hebbian synaptic plasticity ([-1.0, 1.0]) with speaker attribution isolation, continuous geometric taxonomic projection in R^896, and executive working memory with deterministic sub-350-token prompt synthesis.

01
FOUR-TIER COGNITIVE SUBSTRATE Slotted Knowledge • 4D Dynamic Tensor • Continuous Taxonomy • Working Memory Decouples volatile task DAGs from permanent declarative knowledge
02
SIGNED HEBBIAN PLASTICITY Asymmetric Saturation in [-1.0, 1.0] with Speaker Isolation Prevents listener opinion contamination and accelerates contradiction suppression
03
SPECTRAL SHANNON ENTROPY & IHF Dynamically Modulated PageRank Diffusion with Hairball Immunity Analytical hops and Inverse Hub Frequency dampening without arbitrary heuristics
Explore Tenor on GitHub ★ View Paper & Theorems ↓
exorticlabs/tenor // main 4D-EEKG v2
# Clone and install the 4D-EEKG memory engine
$ git clone https://github.com/exorticlabs/tenor.git
$ cd tenor && pip install -e .
$ python
>>> from tenor.orchestrator import MemoryOrchestrator
>>> memory = MemoryOrchestrator(storage_dir="./store")
✓ 4D spatio-temporal tensor initialized (R^896)
✓ signed Hebbian plasticity engine online
→ slotted quotas budget locked to 350 tokens
LICENSE: MIT PYTHON: 3.10+ STATUS: OPEN SOURCE
View Code ↗
FOUNDATIONS REPORT PEER-REVIEW PREPRINT OPEN ACCESS & CODE AVAILABLE

4D-EEKG: A Four-Tier Spatio-Temporal Episodic-Entity Knowledge Graph Architecture with Signed Associative Plasticity and Emergent Continuous Taxonomic Manifolds

Cognitive Systems & Autonomous Agent Architecture Research Group — Exortic Research Lab

Technical Specification & Systems Foundations Report (November 2026)
EXECUTIVE ABSTRACT

Autonomous language agents operating across extended multi-session tasks face fundamental cognitive limitations: flat vector-retrieval architectures suffer from catastrophic semantic drift, lack temporal interval awareness, and fail to distinguish between permanent commitments and transient scratch hypotheses. We formalize 4D-EEKG, a mathematically grounded four-tier memory architecture that unifies structured entity attributes, continuous 4D dynamic associative attention, continuous geometric taxonomic projection, and volatile executive working memory. The architecture introduces signed Hebbian synaptic plasticity with asymmetric diminishing saturation over edge valences in [-1.0, 1.0], strict speaker-attribution isolation, Inverse Hub Frequency (IHF) dampening, and Personalized PageRank diffusion modulated by the spectral Shannon entropy of query seed distributions. Context synthesis is strictly bounded by a deterministic Slotted Quotas allocator enforcing a sub-350-token prompt budget.

[THEOREM 1]

Temporal Slice Invariance

Every entity attribute is stored as a closed interval [t_start, t_end]. Updates strictly supersede without destroying historical lineage, guaranteeing auditability.

[FORMULATION 2]

Bounded Signed Plasticity

Hebbian edge weights saturate asymptotically toward +1.0 for confirmed associations and accelerate toward -1.0 for contradictive invalidations.

[ALGORITHM 3]

Spectral Shannon Diffusion

Information-theoretic entropy of query seed energy analytically modulates PageRank diffusion hops and restart retention without ad-hoc heuristics.

[MANIFOLD 4]

Continuous Taxonomic Field

Replaces brittle discrete relation enums with continuous projections in R^896, featuring online EMA pole adaptation and residual axis discovery.

KEY ARCHITECTURAL FIGURES & EMPIRICAL SCALING
Figure 1: 4D-EEKG Four-Tier Cognitive System Architecture
Figure 1: Four-Tier Cognitive System Topology: Volatile working memory, 4D tensor engine, and long-term slotted knowledge.
Figure 2: Dual-Stream Retrieval Pipeline
Figure 2: Dual-Stream Retrieval Pipeline: Seed extraction, Shannon entropy traversal, and Slotted Quotas context assembly.
Figure 6: Log-Log Physical Memory Scaling Footprint
Figure 6: Memory scaling: Sparse CSR tensors provide a 16,393x compression advantage over dense adjacency matrices at 1M entities.
Read Full Paper (Markdown) ↗ Download Formal Manuscript (.docx) ⭳ View Code on GitHub ★
← Previous: Guidey (SYS-01) Next Drop: Perception Substrate (SYS-03) ⚡