Inherited memory lifted an agent’s office score 11.33 points.

Agents that rewrite their own scaffolding usually restart from zero when the model changes.

A harness tuned for Claude Opus 5.5 scored worse on Gemini 3.8 Flash until a log of past changes let it adapt.

The log records each change’s reason, expected result and outcome, but the model’s weights never change.

How each outlet framed it
Zhidx leans favorable
reports SIM beat baseline in 9 of 10 coding trials (+2.36 avg) and 84.75 vs 73.42 on a cross-model office task

Sources: Zhidx