Remember without retraining.

A picture, not a paper

Think of a spectrum. In hyperdimensional computing, HDC for short, every fact, record, or relationship becomes a wide band of information: a pattern with thousands of positions rather than a single number. Two patterns can be combined into a third. Similar things produce similar bands, so the memory finds them by resemblance. You write a fact down. You do not retrain the world to change it.

Nothing about this is optical. It runs on ordinary processors. The difference is where knowledge lives: as patterns you can read, guard, and edit, rather than as weights buried inside a model.

Language models remain useful. They look, they listen, they write. They should not be the room where the knowledge lives. Our work is the bridge between the two: the neural network in front, the hyperdimensional memory behind it.

The program

Since June 2025 we have filed sixteen applications on hyperdimensional memory: hierarchical encoding, integrity-guarded storage, access control at the signal level, bidirectional querying, latent pattern discovery, high-throughput ingestion, inference over the memory itself, sparse ternary hyperdimensional encoding, and gradient-free knowledge ingestion. The filings use the term “holographic” for the encoding method; the hardware is conventional. Our public writing on ternary computing, neural operating systems, and synthesized intelligence is the argument. The filings are the mechanism. Both are listed on the Scholar page.

How we help

• Translate the pitch, “encode, don’t train,” “geometric memory,” “the thing after transformers,” into a decision a leader can accept or refuse.
• Design the bridge: which knowledge stays in the model, which moves into hyperdimensional memory, and how the two talk to each other in the systems you already run.
• Say when this belongs next to the systems you already have, and when it is a paper dressed as a product. That includes ours.
• Set a fair test: can it remember the right person, admit when it does not know, and change a fact without a ceremony?
• Keep the ownership question in the room: who holds the memory, and who can read it?
• For institutions with a concrete corpus and a concrete question, run a bounded pilot of the memory program under agreement.

A boundary we keep

The advisory work stands on its own. You can hire it without touching our lab, and we will tell you when someone else’s approach is the better fit. Where the memory program is the right answer, that conversation happens in private, after the problem is named.