How NCA Knowledge Encoding Works

The Core Idea

Most AI systems store knowledge as either model weights (billions of parameters baked in at training time) or text chunks (vector databases, RAG pipelines). SAGE does neither. It encodes knowledge into a 256×256 Neural Cellular Automata grid — a living, self-organizing structure where information exists as emergent patterns in cellular dynamics.

This isn't a metaphor. The grid is literally a 2D array of cells, each containing 32 channels of floating-point values. Cells perceive their neighbors and update through tiny neural networks. Knowledge encoding means writing patterns into this grid such that they can be retrieved later through query-driven readout.

Step 1: Text → Tokens

The encoding pipeline starts with standard BPE (Byte Pair Encoding) tokenization. Current experiments use a 1,024-token vocabulary, though this scales to larger vocabularies as grid size increases.

The tokenizer is deliberately small compared to LLM tokenizers (50K–100K tokens). This is intentional — the NCA grid doesn't need to distinguish between "can't" and "cannot" at the token level. The grid's spatial dynamics handle semantic nuance through pattern interaction, not vocabulary size.

Step 2: Spatial Projection (Semantic Hashing)

Each token is mapped to a 2D coordinate on the grid via a learned semantic hash function. This is a small neural network (2 layers, ~500 parameters) that maps token IDs to (x, y) coordinates:

The semantic hash is trained end-to-end with the rest of the system. It learns to place related concepts near each other on the grid — not because we told it to, but because proximity enables the NCA dynamics to form useful associative connections between related knowledge.

Step 3: Gaussian-Weighted Write

Once a token has a grid coordinate, knowledge is written as a Gaussian activation centered on that coordinate. The write isn't a single-cell operation — it's a smooth, overlapping activation pattern:

The Gaussian spread is critical. It means each token's knowledge overlaps with nearby tokens' knowledge. When "coral" writes at (42, 87) and "reefs" writes at (44, 89), their Gaussian activations overlap — and the NCA dynamics in the overlap region create associative patterns that encode the relationship between the concepts.

Confidence Gating

Writes are gated by existing confidence values (channel 18). If a cell already contains high-confidence knowledge, the new write is attenuated:

This prevents new information from overwriting well-established knowledge, while still allowing gradual updates as evidence accumulates.

Step 4: NCA Integration

After writing raw activations to the grid, SAGE runs 10–20 NCA update steps. This is where the magic happens. Each cell:

1. Perceives neighbors via 3×3 Sobel filters (detecting gradients and patterns in nearby cells)
2. Processes perception through a tiny MLP (2 layers, ~50 parameters per cell)
3. Updates state with a stochastic mask (not every cell updates every step — this prevents synchronization artifacts)

These integration steps allow the new knowledge to interact with and merge into existing patterns. The NCA dynamics create associations, resolve conflicts, and build the emergent spatial structure that makes retrieval possible.

Channel Allocation: 24 Shared + 8 Private

Each cell's 32 channels serve specialized roles. This isn't arbitrary — it's a carefully designed allocation that separates concerns:

Channels	Name	Shared?	What They Store
0–3	Structural	✓ Yes	Cell alive/dead state, health metric, connectivity to neighbors, structural role (hub vs leaf)
4–11	Semantic	✓ Yes	Token embedding projections — the core "meaning" of what's stored at this location. 8 dimensions capture the primary semantic axes
12–17	Association	✓ Yes	Cross-cell relationship patterns — co-occurrence signals, causal links, hierarchical relationships. These emerge primarily from NCA dynamics, not direct writes
18–21	Temporal	✓ Yes	When knowledge was written (recency), how many sources confirmed it (confidence), decay rate, update frequency
22–23	Provenance	✓ Yes	Hashed source identity, reputation score of the contributing node. Used for trust-weighted retrieval
24–31	Private	✗ No	Local-only knowledge — personal preferences, PII-adjacent patterns, user-specific context. Never transmitted over the network

Why 24 + 8?

The ratio is empirically determined. 24 shared channels provide enough representational capacity for rich knowledge encoding while keeping diff sizes small (~1KB). 8 private channels give each node room for personalization without affecting the shared knowledge layer. In experiments, reducing shared channels below 20 degraded retrieval quality; increasing private channels beyond 8 didn't improve personalization measurably.

Retrieval: Query → Grid → Context

When SAGE needs to answer a question, it reverses the encoding process — but with important differences:

Step 1: Query Projection

The question is tokenized and each token is projected to grid coordinates using the same semantic hash. This identifies regions of interest on the grid.

Step 2: Attention Readout

For each region of interest, an attention mechanism reads from the grid cells in the neighborhood. The attention weights consider:

• Semantic relevance: How closely the cell's semantic channels (4–11) match the query token's embedding
• Activation strength: Structural channels (0–3) indicate how "alive" and connected the cell is
• Confidence: Temporal channels (18–21) weight recent, well-corroborated knowledge higher
• Association strength: Association channels (12–17) pull in related concepts that weren't directly queried

This last point is crucial. The attention readout doesn't just find exact matches — it follows association patterns to pull in contextually relevant knowledge. When you query "coral reefs" and "threatens," the association channels may activate patterns related to "ocean acidification" and "temperature rise" even though those terms weren't in the query.

Step 3: Context Vector → Text Prefix

The attention readout produces a context vector that is decoded into a natural language prefix. This prefix is prepended to the transformer's input, giving it factual grounding:

Step 4: Confidence Gating

If the readout confidence is below a threshold (default: 0.3), the context is suppressed entirely. The system returns "I don't know" rather than generating a response from low-confidence knowledge. This is SAGE's defense against hallucination — if the grid doesn't have reliable knowledge, it says so.

Why This Isn't a Vector Database

At first glance, NCA knowledge encoding might seem like "just another embedding store." It's not. The differences are fundamental:

The Association Problem

Consider a vector database that has indexed two documents:

1. "Climate change causes ocean acidification and temperature rise"
2. "Coral organisms are sensitive to pH levels and temperature"

If you query "How does climate change affect coral reefs?", a RAG system retrieves both chunks by embedding similarity. But it doesn't understand the connection — it just dumps both chunks into the LLM's context window and hopes the LLM figures it out.

In SAGE's NCA grid, the encoding of "ocean acidification" and "sensitive to pH" creates overlapping activations in nearby grid regions. After NCA integration steps, the cellular dynamics form associative patterns that directly encode the relationship: acidification (from climate change) lowers pH (which coral is sensitive to). The grid computes the association — it doesn't just store the raw facts.

The Fixed-Size Advantage

A vector database grows with data. Index 1 million documents and you need gigabytes of storage plus expensive similarity search infrastructure. The NCA grid is always 128KB. New knowledge is compressed into existing patterns — older, less-confirmed knowledge gradually fades as new patterns overwrite them through confidence gating.

This means SAGE nodes can sync their entire knowledge state in a few hundred milliseconds over a normal internet connection. Try that with a vector database.

Continuous Learning

Vector databases are static indexes. You add documents, you search documents. There's no learning, no integration, no emergence. The NCA grid is a living structure — every new piece of knowledge triggers integration steps that can create new associations, strengthen existing ones, or let outdated knowledge decay. The grid is always changing, always learning.

Concrete Example: Full Pipeline

Let's trace a complete encode-then-retrieve cycle:

Note how "constructed" (the query term) maps near "built" (the stored term) via the semantic hash — even though they're different tokens. The spatial proximity lets the retrieval system find relevant knowledge without exact token matching.

Current Limitations

• Grid capacity: 128KB is small. A single grid can hold perhaps a few hundred well-encoded facts with good retrieval quality. For larger knowledge bases, SAGE uses grid ensembles and hierarchical grids (active research)
• Token vocabulary: 1,024 tokens is limited. Scaling to larger vocabularies requires proportionally larger grids or more sophisticated hashing
• Precision: Lossy compression means exact quotes and precise numbers can degrade. The grid captures gist, not verbatim text
• Training stability: The semantic hash must be trained carefully — a bad hash that doesn't cluster related concepts will produce poor retrieval

Further Reading

• SAGE Whitepaper — Full system architecture including distribution protocol and security model
• NCA Training Results — Signal ratio experiments across corpora and grid sizes
• Reservoir Computing Results — 100% top-5 prediction from frozen NCA dynamics
• Growing Neural Cellular Automata (Mordvintsev et al., 2020) — The foundational NCA paper