CogniRay

CogniRay is an experimental cognitive memory system built upon the principles of Holographic Projection Memory (HPM). It is designed as a structured, differentiable, and geometrically projective memory substrate that supports active inference, directional recall, and adaptive semantic imprinting.

This repository contains:

• Theoretical foundations of Holographic Projection Memory (HPM) and CogniRay (see docs/)
• Minimal working implementation of HPM (read/write logic) (see src/)
• Examples and exploratory notebooks demonstrating usage (see examples/)

---

News

07/2025:

• Public Website Launched: CogniRay is now live at CogniRay.online — featuring a complete interactive introduction to the theory, architecture, and experimental proof of Holographic Projection Memory. Designed as a readable, argument-driven exposition for researchers, engineers, and the cognitively curious. Includes detailed descriptions of projection geometry, update modes, and the semantic behavior of the memory field under real-time adaptation.
• Research Prototype (v0.1.1): Extended reference implementation with verified projection search mechanisms. Includes mem_search_genetic.py (genetic ray synthesis) and mem_search_random_samples_gradient_descent.py (gradient descent over random projections). Confirms navigability of the memory field and supports experimental manipulation of projection vectors in learned memory space. Saved as a separate branch: research-prototype-0-1-1

06/2025:

• Experiment 01: HPM converges rapidly from zero to target projections under exact rays — validating core update mechanics.
• Experiment 02: HPM self-organizes from noise using only inferred projections — stable convergence, semantic drift, no backprop needed.
• Experiment 03: HPM supports classical SGD — matches Delta-Learning performance under noisy rays, confirming architectural compatibility.
• Experiment 04: HPM exhibits semantic drift and adaptive topological divergence under conflicting projection targets — projection geometry adjusts via SGD to preserve memory integrity.
• Experiment 05: Four update modes benchmarked under interference — reflexive update achieves fastest convergence and best memory protection, confirming protective capacity of geometrically filtered updates.
• Experiment 06: Associative dynamics enable spontaneous recovery of erased content — HPM behaves as an attractor system, revealing latent autoassociative memory properties.
• Research Prototype (v0.1.0): First structured release of CogniRay — includes complete reference implementation, formal update modes, and six documented experiments. Demonstrates architectural coherence, functional stability, and practical applicability of Holographic Projection Memory in adaptive reasoning contexts. Saved as a separate branch: research-prototype-0-1-0

---

Key Concepts and Advantages

Geometric Access to Memory

Memory is not accessed by keys or positions but by directional projection rays traversing a latent field. This enables:

• Content-based access via spatial geometry
• Localized attention along semantic trajectories
• View-dependent retrieval, mimicking perception

Differentiable Projection Mechanics

Read and write operations are implemented as smooth, differentiable integrals along rays using soft kernels. This allows:

• Gradient-based learning of both memory and access patterns
• Seamless integration with neural architectures
• Full support for backpropagation and active error correction

Topological Divergence

Conflicting updates do not overwrite but diverge spatially, forming distinct semantic regions. Benefits include:

• Natural separation of contradictory memories
• Controllable avoidance of catastrophic forgetting
• Self-organizing semantic clustering

Inference-Time Plasticity

Memory is not static — it can adapt on the fly by modifying localized regions based on projection errors. This provides:

• Real-time correction of outdated or incomplete knowledge
• On-the-fly learning during inference
• Compatibility with lifelong and self-supervised learning

Multi-Scale Scanning

Hierarchical, multi-resolution projection surfaces allow focus from fine-grained detail to global context, enabling:

• Coarse-to-fine reasoning
• Attention allocation to regions of semantic saliency
• Efficient memory search and reconstruction

Orientation-Aware Retrieval

Projection direction vectors act as semantic filters, encoding preferences for angular alignment:

• Emulates biological orientation selectivity (e.g., V1 cortex)
• Enables disentangled recall based on directional cues
• Supports dynamic, learnable routing mechanisms

Interpretability and Visualizability

Memory operations are explicitly geometric and spatially structured, which ensures:

• Transparent debugging and visualization of memory state
• Modular composability (e.g., stacking beams, splitting projection surfaces)
• Explainability for cognitive systems and symbolic overlays

Compatibility with Discrete and Continuous Models

Though formulated in continuous space, HPM is implemented via voxelized tensors:

• Compatible with GPU-accelerated rasterization and grid processing
• Supports hybrid symbolic-geometric representations
• Bridges low-level neural fields and high-level memory primitives

---

Roadmap

1. Extended Research Release (v0.2)
2. Core CogniRay module for PyTorch
3. Interactive visualization and ray tracing interface

TODO:

• Conduct Experiment 07
• Prepare and conduct Experiment 08
• Prepare and conduct Experiment 09

---

Project & Company

CogniRay is developed by MnemoWare, an independent research initiative building cognitive infrastructures for machine reasoning.

We design memory systems that learn by projecting, adapt by recalling, and scale with understanding.

MnemoWare. Shape your mind.

Project: CogniRay.online
Company: MnemoWare.com (coming soon)

X (ex. Twitter): MnemoWare

Telegram group: CogniRay
Telegram channel: MnemoWare

---

Licensing

Documentation and Theoretical Materials

Licensed under Creative Commons Attribution–NonCommercial–ShareAlike 4.0 International (CC BY-NC-SA 4.0) https://creativecommons.org/licenses/by-nc-sa/4.0/

You are free to read, redistribute, and adapt the documentation for non-commercial purposes, provided that:

• Proper attribution is given (Konstantin Bashinskiy)
• Derivative works are shared under the same license

Code and Implementations

Licensed under the CogniRay Non-Commercial License v1.0 Commercial use is strictly prohibited without prior written permission from the author.

You may use the code for:

• Research
• Education
• Personal non-commercial experimentation

For commercial licensing, please contact: license@cogniray.online

---

Author

Konstantin Bashinskiy (a.k.a. Sombressoul)
AI researcher, memory reconstruction theorist, and occasional code-wrestler.

---

Note: Previous versions contained a typographic error in the author’s surname. This has been corrected (6057f51fc9d2e273cdbe45210c965770e3f84ac7) for legal and attribution clarity.