A mind of waves, weighed in thirds, that remembers

A transformer remembers by attending over everything it has seen, which is why its cost climbs as context grows. We took a different bet: a second-order, resonant recurrence — think of state as a wave evolving under a Verlet-style update — paired with ternary weights in {−1, 0, +1}, so the core does additions, not multiplies. The question that decides whether the bet pays off is blunt: does a fact stick?

Does-it-stick: a genuine, validated GO

🔬Research

As of 2026-07-02 the plastic-mind go/no-go is GO — with caveats we state up front. 🔬 Two experiments:

• Scale. Inject 25 facts into the stream and query them later: recall holds at cosine 0.77, drift-bounded, with capacity around 85–90 (the knee is not reached even at N=60). 🔬
• On real embeddings — a null result, stated as one. Running on real Qwen-0.5B embeddings rather than synthetic vectors gives cosine 0.87 against a random baseline of 0.90. 🔬 That is below chance on this arm: it does not separate from the control, and we no longer present 0.87 as a result. What it does show is no catastrophic forgetting on real, correlated language.

The 25-fact arm above is the measured result we stand behind on this lane. The real-embedding arm is a control we did not beat, and it stays on the page for exactly that reason.

Reshape beats route

🔬Research

A structural result worth its own line: a 22%-width prefix keeps 95% of accuracy — a 4.4× saving — in a regime where a flat network simply collapses. 🔬 Reshaping the representation beats routing around it. This is the mechanism that lets a small, cheap model carry more than its parameter count suggests it should.

The honest negatives — on the page, not in a footnote

🔬Research

• It needs to dream. 🔬 Without a consolidation ("dreaming") pass, the memory is recency-only — it remembers the last thing, not the right thing. Dreaming is what turns the store into a real associative memory.
• Recall is approximate, not verbatim. 🔬 It answers about the fact; it does not quote it back character-for-character. For some uses that is fine; for others it is disqualifying, and we say so.
• An on-disk ternary-7B is a known negative. 🔬 At 7B, on disk, the ternary path runs roughly 7× slower. We keep this failure visible because a lab that hides its negatives forfeits the right to be believed on its positives.

How it works, briefly

State evolves as a damped, driven oscillation — a wave — so information persists through resonance rather than through an attention matrix. Weights are ternary, so the hot loop is add-and-accumulate, not multiply-accumulate: friendly to cheap silicon and to an in-memory analog substrate later. A separate SSM/wave decode path reaches 45× at 16k tokens with a 1.14 MB state — the kind of number that comes from not re-reading the whole context every step.

Why it ladders back

A capable model that runs on hardware you already own — because it does additions, not multiplies, and remembers without a datacentre — is access for everyone, not just the people who can rent a cluster.