Docs

Iterative AI

What iterative AI means, how the experience-to-update loop works and why memory alone is not learning.

Iterative AI is an AI system that changes its future behaviour through evaluated experience.

The phrase is easy to dilute. A chat history is not necessarily iterative. Saving a preference is not necessarily iterative. Retrieving an old answer is not necessarily iterative. A system becomes iterative when experience passes through a loop that can produce a controlled, measurable change.

The loop#

1. Experience — The system acts and produces an outcome. 2. Evidence — The relevant input, output, correction and result are preserved. 3. Interpretation — The system forms a bounded claim about what the evidence means. 4. Guidance — That claim influences a later task where it is relevant. 5. Evaluation — The system or user checks whether the guidance improved the outcome. 6. Update — The claim is confirmed, narrowed, revised or revoked.

The difficult part is not storing experience. It is deciding what experience deserves to become durable.

A rejection may mean “never do this again.” It may also mean “not for this client,” “not today,” “too expensive,” or “wrong for this format.” Treating all feedback as universal creates a system that becomes confident faster than it becomes correct.

Iterative AI therefore requires scope, provenance and uncertainty.

It also requires forgetting. A system that only accumulates eventually becomes crowded with contradictions, outdated intentions and accidental rules. Good learning is partly the ability to know what no longer deserves influence.

Catalyst uses the term iterative AI for systems designed around this full loop rather than systems that simply expose a memory feature.