μLattice: Breaking the Quadratic Memory Barrier in AI

Stream your own document

Paste any text. Define facts as name → value pairs. Mark any to be changed mid-stream. The model reads the stream, stores the facts, and returns the latest value for each.

Your text (becomes the stream)

Your facts (name → value; fill “new value” to overwrite mid-stream)

The store at scale ↗

The same mechanism at scale: store up to 8,192 pairs in one 133,120-token causal pass, overwrite half of them, then query. Measured recall: 1.0000.

pairs 2048

filler/pair 0

changed mid-stream 0%

Left: learned addresses of 400 stored keys from this run (grid scale shown). Right: 64 sampled queries; gold rows marked changed mid-stream were re-written later in the stream — the store must answer with the new value.

Constant cost while memory grows ↗

Each token touches exactly nine memory cells, so per-token work cannot grow with context. Measure it: the curve below runs the model at increasing stream lengths, live.

Blue: measured μs per token — flat is the guarantee. Green: lattice cells, growing with stored content. Dashed red: attention’s O(T²) work, analytic reference.

Inside one cell ↗

Where does a value physically live, and where does the new value go on overwrite? Pick one fact from section 01 (or any name/value). The model’s learned address for that key, the 3×3 cells it touches, and the before/after answers — same cells, value replaced in place by the delta rule.

fact to inspect

Notes

This is a content-addressed associative mixer, not a language model. It stores key→value associations and retrieves them at constant per-token cost. Perfect scores are on the synthetic MQAR distribution (see PAPER.md).

Current demo is 2D (9-cell neighborhood per token). 3D (27 cells) is supported by the kernel math without architectural change — it reduces key density by 1000× and raises the estimated capacity from ~8K to ~495K keys (details).