I built a memory system that lets Claude Haiku (the $1/M-token model) answer questions with 100% accuracy — tying Claude Opus (the $5/M-token model) running with the entire knowledge base in context. Haiku + my memory costs $0.10 per thousand questions. Opus + full context costs $64.04. Same accuracy. 640x cheaper.
Both models score 0% without memory. The facts are synthetic — there's nothing in their training data to fall back on. The memory is the entire difference.
I'm open-sourcing the system today. pip install slate-memory.
What it is
Slate-memory is a one-shot attractor memory. You commit facts by embedding them once. When you query, the system settles into the nearest stored pattern via softmax-weighted feedback — the same math as transformer attention, but used as a lookup table instead of a layer.
