Tenor (4D-EEKG Memory Architecture)
Tenor is a four-tier spatio-temporal memory architecture engineered for autonomous AI agents.
Rather than relying on flat vector databases, naive top-k semantic similarity, or brute-force million-token context windows, Tenor implements an integrated Episodic-Entity Knowledge Graph (4D-EEKG) with signed associative plasticity, continuous taxonomic manifolds, and volatile working memory orchestration.
Read the full scientific report: docs/RESEARCH_PAPER.md
Download Word manuscript: docs/research_paper_4d_eekg_memory_architecture.docx
The Problem: AI Amnesia & The Limits of RAG
Every autonomous agent running in production today hits the exact same walls:
- The KV-Cache Wall: Storing raw 1M-token context buffers consumes over 130 GB of VRAM per session on 70B models. Re-reading entire conversation transcripts on every prompt is computationally unsustainable.
- Semantic Similarity Is Not Truth: Vector search returns text that sounds similar, not what is factually true or currently valid. It cannot handle state mutations (e.g., when a user migrates from Postgres to ClickHouse on Day 4, vector search still returns the Day 1 Postgres recommendation).
- Temporal Blindness: Vector embeddings live in high-dimensional spatial geometry with zero chronological awareness, cause-and-effect reasoning, or dynamic forgetting.
- Graph Explosion ("Hairballs"): Unconstrained symbolic knowledge graphs explode around high-degree hub nodes, diluting graph diffusion queries.
Tenor replaces discrete prompt stuffing with a continuous spatio-temporal substrate.
The Four-Tier System Architecture
┌────────────────────────────────────────────────────────────────────────┐
│ TIER 4: EXECUTIVE WORKING MEMORY │
│ Task DAGs • Tool Execution Telemetry • Slotted Quota Context Packer │
└────────────────────────────────────┬───────────────────────────────────┘
│ Biases & Activates
┌────────────────────────────────────▼───────────────────────────────────┐
│ TIER 2: 4D SPATIO-TEMPORAL TENSOR ENGINE │
│ Signed Hebbian Plasticity ([-1.0, 1.0]) • Speaker Attribution │
│ Inverse Hub Frequency (IHF) • Personalized PageRank Spreading Diff │
└──────────────────┬──────────────────────────────────┬──────────────────┘
│ Anchors │ Projects
┌──────────────────▼───────────────────┐ ┌────────────▼───────────────────┐
│ TIER 1: SLOTTED ENTITY KNOWLEDGE │ │ TIER 3: CONTINUOUS TAXONOMY │
│ TimeSlice Intervals [t_s, t_e] │ │ R^896 Manifold (tanh) │
│ Dual-Path Mutation & Invalidation │ │ Online EMA Residual Discovery │
└──────────────────────────────────────┘ └───────────────────────────────┘
1. Tier 1: Declarative Slotted Entity Knowledge
- Immutable TimeSlice Intervals: Every attribute update is stored as a closed interval $[t_{\text{start}}, t_{\text{end}}]$, ensuring mathematical invariance over updates and full auditability of historical state without data destruction.
- Dual Update Pathways: Supports explicit, non-destructive versioned historical branching alongside verified in-place overwrites.
- Deterministic Dual-Factor Entity Resolution: Matches entities via unified lexical surface boundaries and dense 896-dimensional unit representations.
2. Tier 2: 4D Spatio-Temporal Tensor Engine
- Signed Hebbian Plasticity: Online associative updates bounded strictly in $[-1.0, 1.0]$ with asymmetric diminishing saturation ($+1.0$ for confirmed associations, accelerated suppression toward $-1.0$ for invalidations).
- Speaker Attribution Isolation: Eliminates opinion contamination between agents and users by strictly excluding listeners from associative co-occurrence edge formation.
- Inverse Hub Frequency (IHF): Dampens degree explosion mathematically without arbitrary stopword lists or ad-hoc heuristic thresholds.
- Personalized PageRank Diffusion: Spreading activation modulated dynamically by the Spectral Shannon Entropy of query seed distributions (concentrated factual queries execute sharp local retention; diffuse exploratory queries trigger multi-hop topological diffusion).
3. Tier 3: Emergent Continuous Taxonomic Manifold
- Non-Discrete Classification: Replaces rigid enum labels with continuous semantic vectors ($\tau \in \mathbb{R}^{896}$).
- Adaptive Bipolar Anchors: Centroid poles update online via exponential moving averages (EMA).
- Autonomous Axis Discovery: Extracts emergent conceptual dimensions through orthogonal residual power iteration.
4. Tier 4: Volatile Executive Working Memory
- Task Directed Acyclic Graphs (DAG): Coordinates hierarchical subtask states, progress flags, and active execution barriers.
- Tool Telemetry Feedback: Dynamically links execution failures back to long-term memory for retroactive invalidation.
- Deterministic Slotted Quotas: Assembles prompt context strictly within a sub-350-token budget (50% direct entity facts, 30% topological associations, 20% recent episodic turns).
Architectural Flowchart
Repository Structure
tenor/
├── docs/
│ ├── figures/ # Architectural diagrams & empirical scaling curves
│ │ ├── fig1_system_topology.png
│ │ ├── fig2_retrieval_pipeline.png
│ │ ├── fig3_hebbian_saturation.png
│ │ ├── fig4_temporal_decay.png
│ │ ├── fig5_spectral_entropy.png
│ │ └── fig6_memory_scaling.png
│ ├── RESEARCH_PAPER.md # Full academic preprint in Markdown
│ └── research_paper_4d_eekg_memory_architecture.docx
├── tenor/ # Core Python Architecture Package
│ ├── __init__.py # Package entrypoint & module definitions
│ ├── embedder.py # Dense embedding provider (896-dim unit vectors)
│ ├── llm_client.py # Zero-dependency resilient HTTP inference client
│ ├── llm_manager.py # LLM-in-the-loop state structuring & query parsing
│ ├── nlp_extractor.py # Linguistic filtering & pattern-based mutation detector
│ ├── node_models.py # KnowledgeNode, EntityRegistry & TimeSlice intervals
│ ├── orchestrator.py # Central memory lifecycle & retrieval controller
│ ├── taxonomy.py # Continuous R^896 taxonomic field & residual discovery
│ ├── tensor_engine.py # 4D Spatio-Temporal tensor, Hebbian learning & PageRank
│ ├── turn_extractor.py # Dialogue turn salience parser & temporal binder
│ ├── working_memory.py # Volatile working memory, task DAG & slotted quotas
│ └── workspace_resolver.py # Local directory & persistent state serialization
├── pyproject.toml # Python build configuration
├── LICENSE # MIT License
└── README.md
Quickstart
Installation
git clone https://github.com/exorticlabs/tenor.git
cd tenor
pip install -e .
Basic Usage
from tenor.orchestrator import MemoryOrchestrator
from tenor.working_memory import WorkingMemoryBus
# 1. Initialize the memory substrate
memory = MemoryOrchestrator(
storage_dir="./memory_store",
embedding_dim=896
)
# 2. Record an interaction with strict speaker attribution
memory.process_turn(
turn_id=1,
speaker="user",
text="We are migrating our primary database from Postgres to ClickHouse for analytics.",
timestamp=1730000000.0
)
# 3. Retrieve context with dynamic entropy-modulated diffusion
context = memory.query(
query_text="Which database should we write this analytical migration script for?",
token_budget=350
)
print(context)
# Returns ClickHouse as current ground truth with invalidated historical Postgres TimeSlice
Research & Citation
If you build upon or utilize the 4D-EEKG architecture in your research, please cite:
@article{exortic2026tenor,
title={4D-EEKG: A Four-Tier Spatio-Temporal Episodic-Entity Knowledge Graph Architecture with Signed Associative Plasticity and Emergent Continuous Taxonomic Manifolds},
author={{Exortic Research Lab}},
journal={Exortic Cognitive Systems & Autonomous Agent Foundations Report},
year={2026},
url={https://github.com/exorticlabs/tenor}
}
License
This project is licensed under the MIT License — see the LICENSE file for details.