The Algorithmic Memory: When AI Interrogates African Archives

Archives are no longer just passive storage spaces. In the digital age, they are transforming into operational systems where historical data is filtered, interpreted, and sometimes generated by machines. What happens when an artificial neural network is invited to "imagine" the past? Can computer code become a form of heritage curation in its own right?

These are the dizzying questions that the "Algorithmic Memory" project attempts to answer. This interactive research prototype, designed for our homepage, explores how Artificial Intelligence (AI) reads and reinterprets African photographic archives. Rather than offering a passive video montage, we chose to create a live negotiation between human curation and algorithmic perception, made visible through GLSL shaders and React Three Fiber.

In this article, we dissect the technical, ethical, and curatorial challenges of algorithmic memory applied to African heritage, using our prototype as a case study.

1. From Static Archive to Algorithmic Curation

For decades, the African photographic archive has been frozen in often distant institutions, indexed according to colonial or Eurocentric taxonomies. The massive digitization of these collections has allowed for a gradual reappropriation, but the introduction of Artificial Intelligence fundamentally changes the game.

AI is not a neutral tool. It acts as an algorithmic mediator. When a Machine Learning model processes a database of historical images, it does not just sort them: it extracts patterns, identifies recurrences, and, in the case of generative AI, extrapolates new representations.

This is what researchers call algorithmic censorship or, conversely, systemic hallucination. If a model has been trained on a biased dataset, it will reproduce those biases endlessly, making certain histories invisible or altering specific cultural markers (textiles, architectures, physiognomies). It is to counter this "epistemic drift" that we designed a workflow where humans maintain absolute control over the machine.

2. Project Architecture: Making AI Tangible

The main goal of our prototype was to make the invisible processes of AI legible to a broad audience. We wanted the visitor to feel the work of the machine, its hesitations, and its interpolations, through movement, light, and interaction.

2.1. The Tech Stack

To achieve this level of fluidity in a web browser without requiring plugins, we opted for a hybrid architecture combining modern web development and low-level graphics programming:

Interface and Routing: Next.js 16 (App Router) orchestrated with Tailwind CSS for the choreography of components and texts.
3D Rendering: React Three Fiber (R3F) managing Three.js scenes within an optimized <Canvas>.
Visual Effects: Custom GLSL fragment shaders handling pixel displacement, chromatic distortions, and real-time mask blending.
AI Assets: A combination of Stable Diffusion (for precise control) and Midjourney (for stylization), generating optimized image sequences in WebP format.

2.2. The Reactor Core: GLSL Shaders

Rather than using a simple pre-rendered video—which would have frozen the experience—we developed a shader system so that the archive is constantly re-evaluated by the user's graphics processing unit (GPU).

Here is a simplified excerpt of the code at the heart of our fragment shader, responsible for displacing pixels to reveal AI "hallucinations":

// Pixel-level displacement to simulate the latent process
float noise = fract(sin(uv.y * 50.0 + uTime * 2.0) * 43758.5453);
displacedUv.x += (noise - 0.5) * strength;

In this block:

uv represents the normalized coordinates of each pixel on the screen.
uTime is a global time variable (uniform) injected directly from React, guaranteeing smooth animation.
strength is a dynamic parameter, calibrated according to the target machine's performance (GPU budget) to ensure 60 frames per second (fps) on desktop while gracefully degrading on mobile.

This approach allows the strength of the distortion to be linked to user interaction (mouse movement or page scrolling), creating a sense of "dialogue" with the archive.

3. The Human-in-the-Loop: Preserving Cultural Integrity

Technology is worthless if it crushes meaning. To avoid falling into the visual clichés or cheap exoticism often generated by mainstream AI, we implemented a highly rigorous prompting architecture anchored in archival reality.

3.1. Cultural Prompt Engineering

Every request sent to our AI models follows a strict structure:

Factual Anchoring: We start with authentic descriptors provided by African archivists and historians (exact location, approximate date, event, type of photographic process).
Sensory Cues: We add descriptive layers to guide the visual texture ("cobalt indigo dye," "1960s West African studio lighting," "silver grain").
Cultural Filter: Each prompt is locked by a specific cultural context to prevent the model from drifting into generic stereotypes.

3.2. The Editorial Workflow

The process of creating the visual assets for our homepage was not a one-click affair. It is an iterative loop:

Initial Generation: Stable Diffusion produces candidate images based on our structured prompts.
Human Validation: A curator examines the results, discards cultural aberrations (anachronisms, inappropriate clothing), and annotates interesting compositions.
Enrichment: The selected images go through variations (via Midjourney or Image-to-Image models) to diversify light and depth.
Optimization: The final sequences are color-graded, compressed (texture atlases), and integrated into the GLSL pipeline.

4. Ethical Safeguards and Transparency

Manipulating photographic memory with AI raises immense ethical questions. Are we falsifying history? To address this legitimate concern, our prototype incorporates several safeguards.

First, we maintain a cryptographic provenance log: every frame generated by the AI can be traced back to the archival image (or text) that inspired it.

Second, the intervention of AI is explicitly declared in the user interface. We categorically refuse prompts aimed at distorting specific historical identities or creating deepfakes of tragic events. Every visual experience is signed and acknowledged as an algorithmic interpretation, not as a raw historical document.

5. The Future of Augmented Curation

What we learned by developing The Algorithmic Memory is that the ambiguity of latent space (the transitional state of AI before it finalizes an image) is a powerful metaphor for memory itself. Memory is not a hard drive; it recomposes, it is in motion. Users were more captivated by the fluid transitions between image states than by the finalized images.

By combining the specificity of human curation and the extrapolative power of AI, we are not destroying the archive: we are giving it a new voice. A voice capable of visually filling the silences of colonial history, proposing alternative futures, and engaging a new generation with its heritage in a radically interactive way.

Let's build inclusive cultural platforms together with Ibeji Systems: