TotemicAI

Keeps the observations that preserve how a system changed — turning points, direction, relational structure — so a model can be compressed or updated without losing the history that explains it. Flags when new data breaks what was stored.

—

Architecture class: model-agnostic structural-memory AI. TotemicAI is not itself an LLM, Transformer, or mixture-of-experts model; it can provide structural memory and reacquisition around those or other learning systems.

Validation & data provenance

Primary AML Perturb-seqThree leave-one-patient-out folds; structural memory was compared with equal-size target-stratified random memory.

1% retentionDirection fidelity ≈0.571 vs 0.292 random · relational fidelity 0.376 vs 0.026 · topology recall 0.333 vs 0.142.

All nine primary comparisonsDirection, relation and topology favored structural retention across all three held-out patient fields.

This is structural memory, not representative sampling and not a claim that sparse memory preserves every downstream statistic.