scema-memory

Four memories, one of them mostly unanswerable.

Part of Scematica Omni — an agent runtime that perceives an environment, projects competing futures, ranks them under a stated preference, decides or refuses to, and seals a verifiable record of what it did.

The organising idea across every crate: each layer can say "I don't know", and saying it costs nothing. An agent that cannot express ignorance expresses a number of the right shape instead, and nothing downstream can tell it from a measurement.

Episodic (what happened), semantic (what is believed), procedural (how things are done), and counterfactual (what the branch we declined would have done). Append-only JSONL.

The fourth is the point. A counterfactual records a branch the agent did not take, so its projected utility is known and its realised outcome almost never is — nobody ran it. The rule that falls out:

Unresolved counterfactuals are counted, never scored.

Calibration::mean_abs_error is None, not 0.0, when nothing resolved. Imputing outcomes for untaken branches — from a model, a neighbour, a prior — would mean the system generating its own training signal, and every later decision tuning itself to a fiction.

A corrupt line is skipped and counted, never fatal: one half-written record from a killed process must not make the agent amnesiac, and must not be swallowed either.

Licensed MIT.